Elementary Number Theory

Elementary Number Theory
WISB321
F.Beukers
Department of Mathematics
2012
UU
ELEMENTARY NUMBER THEORY

Frits Beukers
Fall semester 2013
Contents
1 Integers and the Euclidean algorithm
1.1 Integers . . . . . . . . . . . . . . . .
1.2 Greatest common divisors . . . . . .
1.3 Euclidean algorithm for Z . . . . . .
1.4 Fundamental theorem of arithmetic .
1.5 Exercises . . . . . . . . . . . . . . . .
.
.
.
.
.
4
4
7
8
10
12
2 Arithmetic functions
2.1 Denitions, examples . . . . . . . . . . . . . . . . . . . . . . . . .
2.2 Convolution, Mobius inversion . . . . . . . . . . . . . . . . . . . .
2.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
15
15
18
20
3 Residue classes
3.1 Basic properties . . . . . . .
3.2 Chinese remainder theorem
3.3 Invertible residue classes . .
3.4 Periodic decimal expansions
3.5 Exercises . . . . . . . . . . .
.
.
.
.
.
21
21
22
24
27
29
.
.
.
.
.
33
33
38
39
41
44
.
.
.
.
.
.
47
47
48
51
52
55
59
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
4 Primality and factorisation

4.1 Prime tests and compositeness tests . .
4.2 A polynomial time primality test . . .
4.3 Factorisation methods . . . . . . . . .
4.4 The quadratic sieve . . . . . . . . . . .
4.5 Cryptosystems, zero-knowledge proofs .
5 Quadratic reciprocity
5.1 The Legendre symbol . . .
5.2 Quadratic reciprocity . . .
5.3 A group theoretic proof . .
5.4 Applications . . . . . . . .
5.5 Jacobi symbols, computing
5.6 Class numbers . . . . . . .
. . . .
. . . .
. . . .
. . . .
square
. . . .
1
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
. . . .
. . . .
. . . .
. . . .
roots
. . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
CONTENTS
5.7
Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .
6 Dirichlet characters and Gauss sums

6.1 Characters . . . . . . . . . . . . . . .
6.2 Gauss sums, Jacobi sums . . . . . . .
6.3 Applications . . . . . . . . . . . . . .
6.4 Exercises . . . . . . . . . . . . . . . .
7 Sums of squares, Warings problem
7.1 Sums of two squares . . . . . . . .
7.2 Sums of more than two squares . .
7.3 The 15-theorem . . . . . . . . . . .
7.4 Warings problem . . . . . . . . . .
7.5 Exercises . . . . . . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
8 Continued fractions
8.1 Introduction . . . . . . . . . . . . . . . . . .
8.2 Continued fractions for quadratic irrationals
8.3 Pells equation . . . . . . . . . . . . . . . . .
8.4 Archimedess Cattle Problem . . . . . . . .
8.5 Cornacchias algorithm . . . . . . . . . . . .
8.6 Exercises . . . . . . . . . . . . . . . . . . . .
9 Diophantine equations
9.1 General remarks . . . . . .
9.2 Pythagorean triplets . . .
9.3 Fermats equation . . . . .
9.4 Mordells equation . . . .
9.5 The abc-conjecture . . .
9.6 The equation xp + y q = z r
9.7 Mordells conjecture . . .
9.8 Exercises . . . . . . . . . .
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
60
.
.
.
.
62
62
65
67
71
.
.
.
.
.
72
72
74
77
78
81
.
.
.
.
.
.
82
82
85
88
90
91
93
.
.
.
.
.
.
.
.
94
94
94
96
98
100
102
105
105
10 Prime numbers
107
10.1 Introductory remarks . . . . . . . . . . . . . . . . . . . . . . . . . 107
10.2 Elementary methods . . . . . . . . . . . . . . . . . . . . . . . . . 111
10.3 Exercises . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
11 Irrationality and transcendence
11.1 Irrationality . . . . . . . . . . .
11.2 Transcendence . . . . . . . . . .
11.3 Irrationality of (3) . . . . . . .
11.4 Exercises . . . . . . . . . . . . .
F.Beukers, Elementary Number Theory
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
117
117
120
122
124
CONTENTS
12 Solutions to selected problems
125
13 Appendix: Elementary algebra

13.1 Finite abelian groups . . . . .
13.2 Euclidean domains . . . . . .
13.3 Gaussian integers . . . . . . .
13.4 Quaternion integers . . . . . .
13.5 Polynomials . . . . . . . . . .
143
143
146
147
147
150
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
.
Chapter 1
Integers and the Euclidean
algorithm
1.1
Integers
Roughly speaking, number theory is the mathematics of the integers. In any

systematic treatment of the integers we would have to start with the so-called
Peano-axioms for the natural numbers, dene addition, multiplication and ordering on them and then deduce their elementary properties such as the commutative, associatative and distributive properties. However, because most students
are very familiar with the usual rules of manipulation of integers, we prefer to
shortcut this axiomatic approach. Instead we simply formulate the basic rules
which form the basis of our course. After all, we like to get as quickly to the
parts which make number theory such a beautiful branch of mathematics.
We start with the natural numbers
N : 1, 2, 3, 4, 5, . . .
On N we have an addition (+) and multiplication ( or ) law and a well-ordering
(>, <, , ). By a well-ordering we mean that
1. For any distinct a, b N we have either a > b or a < b.
2. From a < b and b < c follows a < c
3. There is a smallest element, namely 1. So a 1 for all a N.
We shall assume that we are all familiar with the usual rules of addition and
multiplication.
1. For all a, b N: a + b = b + a and ab = ba (commutativity of addition and
multiplication).
4
1.1. INTEGERS
2. For all a, b, c N: (a + b) + c = a + (b + c) and (ab)c = a(bc) (associativity

of addition and multiplication)
3. For all a, b, c N: a(b + c) = ab + ac (distributive law).
4. For all a N: 1 a = a.
5. For all a N: a + 1 > a.
6. For all a, b, c N: b > c a + b > a + c and b c ab ac.
7. For all a, b N: a > b there exists c N such that a = b + c.
We shall also use the following fact.
Theorem 1.1.1 Every non-empty subset of N has a smallest element.
Then there is the principle of induction.
Theorem 1.1.2 Let S N and suppose that
1. 1 S
2. For all a N: a S a + 1 S
Then S = N.
Theorem 1.1.2 follows from Theorem 1.1.1 in the following way. Let S be as
in Theorem 1.1.2 and consider the complement S c . This set is either empty, in
which case Theorem 1.1.2 is proven, or S c is non-empty. Let us assume the latter.
Theorem 1.1.1 states that S c has a smallest element, which we denote by a. If
a = 1, then a S, violating the rst condition of Theorem 1.1.2. If a > 1 then
a 1 S c . Hence a 1 S and a S, violating the second condition. We
conclude that S c is empty, hence S = N.
We call a subset S N nite if there exists m N such that s < m for all s S.
There are two concepts which partially invert addition and multiplication.
1. Subtraction Let a, b N and a > b. Then there exists a unique c N
such that a = b + c. We call c the dierence between a and b. Notation;
a b.
2. Divisibility We say that the natural number b divides a if there exists
c N such that a = bc. Notation: b|a, and b is called a divisor of a.
There are many well-known, almost obvious, properties which are not mentioned
in the above rules, but which nevertheless follow in a more or less straightforward
way. As an exercise you might try to prove the following properties.
CHAPTER 1. INTEGERS AND THE EUCLIDEAN ALGORITHM

1. For all a, b, c N: a + b = a + c b = c
2. For all a, b, c N: ab = ac b = c.
3. For all a, b, d N: d|a, d|b d|(a + b)
4. Any a N has nitely many divisors.
5. Any nite set of natural numbers has a biggest element.
Although divison of one number by another usually fails we do have the concept
of division with remainder.
Theorem 1.1.3 (Euclid) Let a, b N with a > b. Then either b|a or there
exist q, r N such that
a = bq + r,
r < b.
Moreover, q, r are uniquely determined by these (in)equalities.
Proof. Suppose b does not divide a. Consider all multiples of b which are less
than a. This is a non-empty set, since b < a. Choose the largest multiple and
call it bq. Then clearly a bq < b. Conversely, if we have a multiple qb such that
a bq 1 that have only the trivial divisors 1, p. Here are the
rst few:
2, 3, 5, 7, 11, 13, 17, 19, 23, 29, 31, . . .
Most of us have heard about them at a very early age. We also learnt that there
are innitely many of them and that every integer can be written in a unique way
as a product of primes. These are properties that are not mentioned in our rules.
So one has to prove them, which turns out to be not entirely trivial. This is the
beginning of number theory and we will take these proofs up in this chapter.
In the history of arithmetic the number 0 was introduced after the natural
numbers as the symbol with properties 0 a = 0 for all a and a + 0 = a
for all a. Then came the negative numbers -1,-2,-3,. . . with the property that
1 + 1 = 0, 2 + 2 = 0, . . .. Their rules of addition and multiplication are
uniquely determined if we insist that these rules obey the commutative, associate
and distributive laws of addition and multiplication. Including the infamous
minus times minus is plus which causes so many high school children great
headaches. Also in the history of mathematics we see that negative numbers
and their arithmetic were only generally accepted at a surprisingly late age, the
beginning of the 19th century.
1.2. GREATEST COMMON DIVISORS
From now on we will assume that we have gone through all these formal introductions and we are ready to work with the set of integers Z, which consists of
the natural numbers, their opposites and the number 0.
The main role of Z is to have extended N to a system in which the operation
of subtraction is well-dened for any two elements. One may proceed further by
extending Z to a system in which also element (= 0) divides any other. The
smallest such system is well-known: Q, the set of rational numbers. At several
occasion they will also play an important role.
1.2
Greatest common divisors
Definition 1.2.1 Let a1 , . . . , an Z, not all zero. The greatest common divisor
of a1 , . . . , an is the largest natural number d which divides all ai
Notation: (a1 , . . . , an ) or gcd(a1 , . . . , an ).
Definition 1.2.2 Two numbers a, b Z, not both zero, are called relatively
prime if gcd(a, b) = 1.
Theorem 1.2.3 Let ai Z (i = 1, . . . , n) not all zero. Let d = gcd(a1 , . . . , an ).
Then there exist t1 , . . . , tn Z such that d = a1 t1 + + an tn
Proof. Consider the set
S = {a1 x1 + + an xn | x1 , , xn Z}
and choose its smallest positive element. Call it s. We assert that d = s. First
note that every element of S is a multiple of s. Namely, choose x S arbitrary.
Then x ls S for every l Z. In particular, x [x/s]s S. Moreover,
0 x [x/s]s < s. Because s is the smallest positive element in S , we have
necessarily x [x/s]s = 0 and hence s|x. In particular, s divides ai S for every
i. So s is a common divisor of the ai and hence s d. On the other hand we
know that s = a1 t1 + + an tn for suitable t1 , , tn . From d|ai i follows that
d|s. Hence d s. Thus we conclude d = s.
2
Corollary 1.2.4 Assume that the numbers a, b are not both zero and that a1 , . . . , an
are not all zero.
i. Every common divisor of a1 , , an divides gcd(a1 , . . . , an ).
Proof: This follows from Theorem 1.2.3. There exist integers t1 , . . . , tn such
that gcd(a1 , . . . , an ) = a1 t1 + + an tn . Hence every common divisor of
a1 , . . . , an divides their greatest common divisor.

ii. Suppose gcd(a, b) = 1. Then a|bc a|c.
Proof: x, y Z : 1 = ax + by. So, c = acx + bcy. The terms on the right
are divisible by a and consequently, a|c.
iii. Let p be a prime. Then p|bc p|b or p|c.
Proof: Suppose for example that p/|b, hence gcd(b, p) = 1. From (ii.) we
infer p|c.
iv. Let p be a prime and suppose p|a1 a2 an . Then i such that p|ai .
Proof Use (iii.) and induction on n.
v. Suppose gcd(a, b) = 1. then b|c and a|c ab|c.
Proof: x, y Z : 1 = ax + by. So, c = acx + bcy. Because both terms on
the right are divisible by ab we also have ab|c.
vi. gcd(a1 , . . . , an ) = gcd(a1 , . . . , an2 , (an1 , an )).
Proof: Every common divisor of an1 and an is a divisor of (an1 , an ) and
conversely (see i.). So, the sets a1 , . . . , an and a1 , . . . , an2 , (an1 , an ) have
the same common divisors. In particular they have the same gcd.
vii. gcd(a, b) = d gcd(a/d, b/d) = 1.
Proof: There exist x, y Z such that ax+by = d. Hence (a/d)x+(b/d)y = 1
and so any common divisor of a/d and b/d divides 1, i.e. d = 1.
viii. gcd(a, b) = 1 there exist x, y Z : ax + by = 1.
1.3
Euclidean algorithm for Z
In this section we describe a classical but very ecient algorithm to determine the
gcd of two integers a, b and the linear combination of a, b which yields gcd(a, b).
First an example. Suppose we want to determine (654321,123456) . The basic
idea is that gcd(a, b) = gcd(a rb, b) for all r Z. By repeatedly subtracting
the smallest term from the largest, we can see to it that the maximum of the
numbers between the gcd brackets decreases. In this way we get
=
=
=
=
=
=
gcd(654321, 123456) = gcd(654321 5 123456, 123456) = gcd(37041, 123456)

gcd(37041, 123456 3 gcd(37041) = gcd(37041, 12333)
gcd(37041 3 gcd(12333, 12333) = gcd(42, 12333)
gcd(42, 12333 293 42) = gcd(42, 27)
gcd(42 27, 27) = gcd(15, 27)
gcd(15, 27 15) = gcd(15, 12)
gcd(15 12, 12) = gcd(3, 12) = gcd(3, 0) = 3
1.3. EUCLIDEAN ALGORITHM FOR Z
So we see that our greatest common divisor is 3. We also know that there exist
integers x, y such that 3 = 654321x + 123456y. To obtain such numbers we have
to work in a more schematic way where we have put a = 654321, b = 123456,
654321 = 5 123456 + 37041

123456 = 3 37041 + 12333
37041 = 3 12333 + 42
12333 = 293 42 + 27
42 = 1 27 + 15
27 = 1 15 + 12
15 = 1 12 + 3
12 = 4 3 + 0
654321 = 1a + 0b
123456 = 0a + 1b
37041 = 1a 5b
12333 = 3a + 16b
42 = 10a 53b
27 = 2933a + 15545b
15 = 2943a 15598b
12 = 5876a + 31143b
3 = 8819a 46741b
0 = 123456a + 654321b
In the left hand column we have rewritten the subtractions. In the righthand
column we have written all remainders as linear combinations of a = 654321 and
b = 123456.
In general, write r1 = a en r0 = b and inductively determine ri+1 for i 0 by
ri+1 = ri1 [ri1 /ri ]ri until rk+1 = 0 for some k. Because r0 > r1 > r2 > 0
such a k occurs. We claim that rk = gcd(a, b). This follows from the following
observation, gcd(ri , ri1 ) = gcd(ri , ri1 [ri1 /ri ]ri ) = gcd(ri , ri+1 ) = gcd(ri+1 , ri )
for all i. Hence gcd(b, a) = gcd(r0 , r1 ) = gcd(r1 , r0 ) = = gcd(rk , rk+1 ) =
gcd(rk , 0) = rk .
In our example we see in the right hand column a way to write 3 as a linear
combination of a = 654321 and b = 123456. The idea is to start with 654321 =
1 a + 0 b and b = 0 a + 1 b and combine these linearly as prescribed by
the euclidean algorithm. We obtain consecutively the ri as linear combination of
654321 and 123456 until
3 = 8819 654321 46741 123456.
In general, let ri (i 1) be as above and suppose rk = 0 and ri = 0 (i < k).
Let
x1 = 1, y1 = 0, x0 = 0, y0 = 1
and inductively,
xi+1 = xi1 [ri1 /ri ] xi ,
yi+1 = yi1 [ri1 /ri ] yi
(i 0).
Then we can show by induction on i that ri = axi + byi for all i 1. In

particular we have gcd(a, b) = rk = axk + byk .
The proof of the termination of the euclidean algorithm is based on the fact that
the remainders r1 , r2 , . . . form a strictly decreasing sequence of positive numbers.
10
In principle this implies that the number of steps required in the euclidean algorithm can be as large as the number a or b itself. This would be very bad if
the numbers a, b would have more than 15 digits, say. However, the very strong
point of the euclidean algorithm is that the number of steps required is very
small compared to the size of the starting numbers a, b. For example, 3100 1
and 2100 1 are numbers with 48 and 31 digits respectively. Nevertheless the
euclidean algorithm applied to them takes only 54 steps. Incidently, the gcd is
138875 in this case. All this is quantied in the following theorem.
Theorem 1.3.1 Let a, b N with a > b and apply the euclidean algorithm to
them. Use the same notations as above and suppose that rk is the last non-zero
remainder in the algorithm. Then
k<2
log a
.
log 2
In other words, the number of steps in the euclidean algorithm is bounded by a

linear function in the number of digits of a.
Proof. First we make an important obervation. Apply the rst step of the
euclidean algorithm to obtain q, r Z0 such that a = bq + r with 0 r < b.
Then we assert that r < a/2. Indeed, if b > a/2 then necessarily q = 1 and
r = a b < a a/2 = a/2. If b a/2 we have automatically r < b a/2.
This observation can also be applied to any step ri1 = qi ri +ri+1 in the euclidean
algorithm. Hence ri+1 < ri1 /2 for every i. By induction we now nd r1 <
a/2, r3 < a/4, r5 < a/8, . . . and in general rl < a/2(l+1)/2 for every odd l. When
l is even we observe that l 1 is odd, so rl < rl1 < a/2l/2 . So we nd for all
l N that rl < a/2l/2 . In particular, rk < a/2k/2 . Using the lower bound rk 1
our theorem now follows.
2
In the exerciseswe will show that the better, and optimal, estimate k < log(a)/ log()
with = (1 + 5)/2 is possible.
1.4
Fundamental theorem of arithmetic
Definition 1.4.1 A prime number is a natural number larger than 1, which has
only 1 and itself as positive divisor.
Usually the following theorem is taken for granted since it is basically taught at
elementary school. However, its proof requires some work and is an application
of Corollary 1.2.4(iv).
Theorem 1.4.2 (Fundamental theorem of arithmetic) Any integer larger
than 1 can be written uniquely, up to ordering of factors, as the product of prime
numbers.
1.4. FUNDAMENTAL THEOREM OF ARITHMETIC
11
Proof. First we show that any n N>1 can be written as a product of primes.
We do this by induction on n. For n = 2 it is obvious. Suppose n > 2 and
suppose we proved our assertion for all numbers below n. If n is prime we are
done. Suppose n is not prime. Then n = n1 n2 , where 1 < n1 , n2 < n. By our
induction hypothesis n1 and n2 can be written as a product of primes. Thus n is
a product of primes.
We now prove our theorem. Let n be the smallest number having two dierent
prime factorisations. Write
n = p1 p2 pr = q 1 q 2 q s ,
where pi , qj are all primes. Notice that p1 |q1 qs . Hence according to Corollary 1.2.4(iv), p1 |qt for some t. Since qt is prime we have p1 = qt . Dividing out
the common prime p1 = qt on both factorisations we conclude that n/p1 has also
two dierent factorisations, contradicting the minimality of n.
2
The proof of the above theorem relies on Corollary 1.2.4 of Theorem 1.2.3. The
latter theorem relies on the fact that Z is a euclidean domain. In Theorem 13.2.3
we have given the analogue of Theorem 1.2.3 for general euclidean domains and in
principle it is possible to prove a unique prime factorisation property for arbitrary
commutative euclidean domains. We have not done this here but instead refer to
standard books on algebra.
Definition 1.4.3 Let a1 , . . . , an Z be non-zero. The lowest common multiple
of a1 , . . . , an is the smallest positive common multiple of a1 , . . . , an .
Notation: lcm(a1 , . . . , an ) or [a1 , . . . , an ].
The following lemma is a straightforward application of Theorem 1.4.2,
Lemma 1.4.4 Let a, b N be non-zero. Write
a = pk11 pk22 pkr r ,
mr
1 m2
b = pm
1 p 2 pr ,
where p1 , . . . , pr are distinct primes and 0 ki , mj (i, j = 1, . . . r). Then,

min(k1 ,m1 )
gcd(a, b) = p1
r ,mr )
pmin(k
r
and
max(k1 ,m1 )
lcm(a, b) = p1
r ,mr )
pmax(k
.
r
Proof. Excercise.
Another fact which is usually taken for granted is that there are innitely many
primes. However, since it does not occur in our Axioms, we must give a proof.
Theorem 1.4.5 (Euclid) There exist innitely many primes.
Suppose that there exist only nitely many primes p1 , . . . , pn . Consider the number N = p1 pn + 1. Let P be a prime divisor of N . Then P = pi for some i
and pi |(p1 pn + 1) implies that pi |1. This is a contradiction, hence there exist
innitely many primes.
2
12
1.5
Exercises
Exercise 1.5.1 Prove, n N :

induction on n)
13 + 23 + + n3 = (1 + 2 + + n)2 (use
Exercise 1.5.2 Prove, 2n 1 is prime n is prime.

(The converse is not true, as shown by 89|211 1.)
Exercise 1.5.3 Prove, 2n + 1 is prime n = 2k , k 0.
(Fermat thought that the converse is also true. However, Euler disproved this by
5
showing that 641|22 + 1.)
Exercise 1.5.4 (**) Prove, n4 + 4n is prime n = 1.
Exercise 1.5.5 Prove that there exist innitely many primes p of the form p
1(mod 4). Hint: use a variant of Euclids proof.
Exercise 1.5.6 Prove that there exist innitely many primes p such that p + 2
is not a prime.
Exercise 1.5.7 (*)Let x, y N. Prove that x + y 2 and y + x2 cannot be both a
square.
Exercise 1.5.8 Prove that n2 1(mod 8) for any odd n N.
Exercise 1.5.9 Determine d = (4655, 12075) and determine x, y Z such that
d = 4655x + 12075y.
Exercise 1.5.10 Let a, b be relatively prime positive integers and c Z and
consider the equation
ax + by = c
in the unknowns x, y Z.
1. Show that there exists a solution.
2. Let x0 , y0 be a solution. Show that the full solution set is given by
x = x0 + bt,
y = y0 at
where t runs through Z.

3. Solve the equation 23x + 13y = 3 in x, y Z completely
Exercise 1.5.11 Let a, b be any positive integers and c Z and consider the
equation
ax + by = c
in the unknowns x, y Z.
1.5. EXERCISES
13
1. Show that the equation has a solution if and only if gcd(a, b) divides c.
2. Describe the full solution set of the equation.
3. Solve the equation 105x + 121y = 3 in x, y Z completely.
Exercise 1.5.12 Let a, b N and suppose a > b. Choose r, q Z such that
a = bq + r and 0 r < b.
a) Show that r < a/2.
Consider now the euclidean algoritm with the notations from the course notes.
Let rk be the last non-zero remainder.
b) Prove that k < log a/ log 2.

Using the following steps we nd an even better estimate for k,
c) Prove by induction on n = k, k 1, . . . , 1 that rn ( 12 (1 + 5))kn .
d) Prove that k < log a/ log( 12 (1 + 5)).

Exercise 1.5.13 To any pair a, b N there exist q, r Z such that a = qb + r
and |r| b/2. Based on this observation we can consider an alternative euclidean
algorithm.
a) Carry out this algorithm for a = 12075 and b = 4655.
b) Suppose a > b and let rk be the last non-zero remainder. Show that k <
log a/ log 2.
Exercise 1.5.14 Let a, m, n N and a 2. Prove that (am 1, an 1) =
a(m,n) 1.
Exercise 1.5.15 (*) Let be a positive rational number. Choose the smallest
integer N0 so that 1 = 1/N0 0. Next choose the smallest integer N1 > N0
such that 2 = 1 1/N1 0. Then choose N2 , etcetera. Show that there exists
an index k such that k = 0. (Hint: First consider the case when 0 < 1).
Conclude that every positive rational number can be written in the form
1
1
1
=
+
+ +
N0 N1
Nk1
where N0 , N1 , . . . , Nk1 are distinct positive integers.
Exercise 1.5.16 In how many zeros does the number 123! end?
Exercise 1.5.17 Let p be prime and a N. Dene vp (a) = max{k Z0 | pk |a}.
a) Prove that
[ ] [ ] [ ]
n
n
n
vp (n!) =
+ 2 + 3 + n.
p
p
p
b) Let n N and write n in base p, i.e. write n = n0 + n1 p + n2 p2 + + nk pk
with 0 ni < p for all i. Let q be the number of ni such that ni p/2. Prove
that
(( ))
2n
vp
q.
n
14
Exercise 1.5.18 Let b, c N and suppose (b, c) = 1. Prove,

a) For all a Z: (a, bc) = (a, b)(a, c).
b) For all x, y Z: (bx + cy, bc) = (c, x)(b, y).
Exercise 1.5.19 (**) Let a, b, c Z not all zero such that gcd(a, b, c) = 1. Prove
that there exists k Z such that gcd(a kc, b) = 1.
Exercise 1.5.20 Prove or disprove: There are innitely many primes p such
that p + 2 and p + 4 are also prime.
Chapter 2
Arithmetic functions
2.1
Definitions, examples
In number theory we very often encounter functions which assume certain values
on N. Well-known examples are,
i. The unit function e dened by e(1) = 1 and e(n) = 0 for all n > 1.
ii. The identity function E dened by E(n) = 1 for all n N.
iii. The power functions Ik dened by Ik (n) = nk for all n N. In particular,
E = I0 .
iv. The number of prime divisors of n, denoted by (n).
v. The number of distinct prime divisors of n, denoted by (n).
vi. The divisor sums l dened by
l (n) =
dl .
d|n
In particular we write = 1 , the sum of divisor and d = 0 , the number

of divisors.
vii. The Euler -function or totient function
(n) = #{d N|gcd(d, n) = 1 and d n}.
viii. Ramanujans -function (n) dened by
(n)x = x
n=1
k=1
15
(1 xk )24 .
16
CHAPTER 2. ARITHMETIC FUNCTIONS
ix. The sums of squares function rd (n) given by the number of solutions
x1 , . . . , xd to n = x21 + + x2d .
In general,
Definition 2.1.1 An arithmetic function is a function f : N C.
Of course this is a very broad concept. Many arithmetic functions which occur
naturally have interesting additional properties. One of them is the multiplicative
property.
Definition 2.1.2 Let f be an arithmetic function with f (1) = 1. Then f is
called multiplicative if f (mn) = f (m)f (n) for all m, n with (m, n) = 1 and
strongly multiplicative if f (mn) = f (m)f (n) for all m, n.
It is trivial to see that examples e, E, Il , 2 are strongly multiplicative and that
2 is multiplicative. In this chapter we will show that l and are multiplicative.
The multiplicative property of Ramanujans is a deep fact based on properties
of so-called modular forms. It was rst proved by Mordell in 1917. As an aside we
also mention the remarkable congruence (n) 11 (n)(mod 691) for all n N.
The multiplicative property of r2 (n)/4 will be proved in the chapter on sums of
squares.
Theorem 2.1.3
i. l (n) is a multiplicative function.
ii. Let n = pk11 pkr r . Then

l (n) =
pl(ki +1) 1
i
pl 1
Proof. Part i. The proof is based on the fact that if d|mn and (m, n) = 1 then
d can be written uniquely in the form d = d1 d2 where d1 |m, d2 |n. In particular
d1 = (m, d), d2 = (n, d). We have
(d1 d2 )l
dl =
l (mn) =
d1 |m,d2 |n
d|mn
dl1
dl2
d1 |m
d2 |n
= l (m)l (n)
Part ii. It suces to show that l (pk ) = (pl(k+1) 1)/(pl 1) for any prime power
pk . The statement then follows from the multiplicative property of l . Note that,
l (pk ) = 1 + pl + p2l + + pkl =
pl(k+1) 1
.
pl 1
2
A very ancient problem is that of perfect numbers.

2.1. DEFINITIONS, EXAMPLES
17
Definition 2.1.4 A perfect number is a number n N which is equal to the sum

of its divisors less than n. Stated alternatively, n is perfect if (n) = 2n.
Examples of perfect numbers are 6, 28, 496, 8128, 33550336, ... . It is not known
whether there are innitely many. It is also not known if there exist odd perfect
numbers. If they do, they must be at least 10300 . There is an internet search for
the rst odd perfect number (if it exists), see /www.oddperfect.org.
For even perfect numbers there exists a characterisation given by Euclid and
Euler.
Theorem 2.1.5 Let n be even. Then n is perfect if and only if it has the form
n = 2k1 (2k 1) with 2k 1 prime.
Proof. Suppose n = 2p1 (2p 1) with 2p 1 prime. Then it is straightforward
to check that (n) = 2n.
Suppose that n is perfect. Write n = 2k1 m, where m is odd and k 2. Then,
(n) = (2k1 m) = (2k1 )(m) = (2k 1)(m).
On the other hand, n is perfect, so (n) = 2n, which implies that 2k m = (2k
1)(m). Hence
m
(m) = m + k
.
2 1
Since (m) is integral, 2k 1 must divide m. Since k 2 we see that m and
m/(2k 1) are distinct divisors of m. Moreover, they must be the only divisors
since their sum is already (m). This implies that m is prime and m/(2k 1) = 1,
i.e m = 2k 1 is prime.
2
Numbers of the form 2m 1 are called Mersenne numbers . If 2m 1 is prime
we call it a Mersenne prime. It is an easy excercise to show that 2m 1 prime
m is prime. Presumably there exist innitely many Mersenne primes but this
is not proved yet. The known values of m for which 2m 1 is prime, are
n=
2, 3, 5, 7, 13, 17, 19, 31, 61, 89, 107, 127, 521, 607, 1279, 2203,
2281, 3217, 4253, 4423, 9689, 9941, 11213, 19937, 21701,
23209, 44497, 86243, 110503, 132049, 216091, 756839, 859433,
1257787, 1398269, 2976221, 3021377, 6972593, 13466917,
20996011, 24036583, 25964951, 30402457, 32582657,
37156667, 42643801, 42643801, 43112609
At the moment (August 2008) 243112609 1 is the largest known prime. For the
latest news on search activities see: www.mersenne.org.
An equally classical subject is that of amicable numbers that is, pairs of numbers
m, n such that n is the sum of all divisors of m less than m and vice versa. In
18
other words, m+n = (n) and n+m = (m). The pair 220, 284 was known to the
ancient Greeks. Euler discovered some 60 pairs (for example 11498355, 12024045)
and later computer searches yielded several thousands of new pairs, some of which
are extremely large.
2.2
Convolution, M
obius inversion
Definition 2.2.1 Let f and g be two arithmetic functions. Their convolution

product f g is dened by
(f g)(n) =
f (d)g(n/d).
d|n
It is an easy excercise to verify that the convolution product is commutative and

associative. Moreover, f = e f for any f . Hence arithmetic functions form a
semigroup under convolution.
Theorem 2.2.2 The convolution product of two multiplicative functions is again
multiplicative.
Proof. Let f, g be two multiplicative functions. We have trivially that (f
g)(1) = f (1)g(1). For any m, n N with (m, n) = 1 we have
( mn )
f (d)g
(f g)(mn) =
d
d|mn
(
)
mn
f (d1 d2 )g
=
d1 d2
d1 |m d2 |n
(
(
)
)
m
n
f (d1 )g
=
f (d2 )g
d1
d2
d1 |m
d2 |n
= (f g)(m)(f g)(n).
2
Notice that for example l = E Il . The multiplicative property of l follows
directly from the mulitiplicativity of E and Il . We now introduce an important
multiplicative function.
Definition 2.2.3 The Mobius function (n) is dened by (1) = 1, (n) = 0 if
n is divisible by a square > 1 and (p1 pt ) = (1)t for any product of distinct
primes p1 , . . . , pt .
Notice that is a multiplicative function. Its importance lies in the following
theorem.

2.2. CONVOLUTION, MOBIUS
INVERSION
19
Theorem 2.2.4 (M
obius inversion) Let f be an arithmetic function and let
F be dened by
F (n) =
f (d).
d|n
Then, for any n N,

f (n) =
F (d)
d|n
(n)
d
Proof. More cryptically we have F = E f and we must prove that f = F .

It suces to show that e = E since this implies F = E f = e f = f .
The function E is again multiplicative, hence it suces to compute E at
prime powers pk and show that it is zero there. Observe,
(E )(pk ) =
(d) = (1) + (p) + + (pk )
d|pk
= 1 1 + 0 + + 0 = 0.
2
Theorem 2.2.5 Let be the Euler -function. Then,
i.
n=
(d),
n 1.
d|n
ii. is multiplicative.
iii.
(n) = n
(
p|n
1
1
p
)
.
Proof. Part i). Fix n N. Let d|n and let Vd be the set of all m n such
that (m, n) = d. After dividing everything by d we see
that |Vd | = (n/d).
Since
{1, . . . , n} = d|n Vd is a disjoint union we nd that n = d|n (n/d) = d|n (d),
as asserted.
Part ii). We have seen in part i) that I1 = E . Hence, by Mobius inversion,
= I1 . Multiplicativity of automatically follows from the multiplicativity
of and I1 .
Part iii). Because of the multiplicativity of it suces to show that (pk ) =
pk (1 1/p). This follows from (pk ) = (I1 )(pk ) = pk pk1 = pk (1 p1 ). 2
20
2.3
Exercises
Exercise 2.3.1 Prove that the sum of the inverses of all divisors of a perfect
number is two.
Exercise 2.3.2 Prove that an odd perfect number must contain at least three
distinct primes.
(It is known that there must be at least six distinct primes).
Exercise 2.3.3 Describe the multiplicative functions which arise from the following convolution products (the symbol 2 stands for the multiplicative function
2(n) ).
Ik Ik
I1
2 2
Exercise 2.3.4 Prove that there exist innitely many n such that (n) + (n +
1) = 0.
Exercise 2.3.5 Prove that there exist innitely many n such that (n) + (n +
1) = 1.
Exercise 2.3.6 Let F be the multiplicative function such that F (n) = 1 if n
is a square, and 0 otherwise. Determine a multiplicative function f such that
f E = F.
Exercise 2.3.7 Let f be a multiplicative function. Prove that there exists a
multiplicative function g such that f g = e (hence the multiplicative functions
form a group with respect to the convolution product).
Exercise 2.3.8 Prove that for every n N
2
0 (k) .
0 (k)3 =
k|n
k|n
Exercise 2.3.9 (**) Show that there exist innitely many n such that (n) is a
square.
Chapter 3
Residue classes
3.1
Basic properties
Definition 3.1.1 Let m N. Two integers a and b are called congruent modulo
m if m|a b.
Notation: a b(mod m)
Definition 3.1.2 The residue class a mod m is dened to be the set {b Z | b
a(mod m)}. The set of residue classes modulo m is denoted by Z/mZ.
The following properties are more or less straightforward,
Remark 3.1.3 Let a, b, c, d Z and m, n N.
i. a a + rm(mod m) r Z.
ii. There are m residue classes modulo m.
iii. a b(mod m) and n|m a b(mod n).
iv. a b(mod m) and c d(mod m) a + c b + d(mod m), ac
bd(mod m).
v. Let P (x) Z[x]. Then a b(mod m) P (a) P (b)(mod m).
vi. The residue classes modulo m form a commutative ring with 1-element.
Its additive group is cyclic and generated by 1 mod m. If m is prime then
Z/mZ is a nite eld. If m is composite then Z/mZ has divisors of zero.
Definition 3.1.4 Let m N. A residue class a mod m is called invertible if
x mod m such that ax 1(mod m). The set of invertible residue classes is
denoted by (Z/mZ) .
21
22
CHAPTER 3. RESIDUE CLASSES
Notice that (Z/mZ) is nothing but the unit group of Z/mZ. In particular the
inverse x mod m of a mod m is uniquely determined modulo m.
Theorem 3.1.5 Let a Z en m N. Then,
a mod m is invertble (a, m) = 1.
Proof. The residue class a mod m is invertible x : ax 1(mod m)
x, y : ax + my = 1 (a, m) = 1.
2
Notice that the proof of the latter theorem also shows that we can compute
the inverse of a residue class via the euclidean algorithm. For example, solve
331x 15(mod 782). The euclidean algorithm shows that the g.c.d. of 331 and
782 is 1 and 1 = 189 331 + 80 782. Hence 189 mod 782 is the inverse of
331 mod 782. To solve our question, multiply on both sides with 189 to obtain
x 293(mod 782).
3.2
Chinese remainder theorem
Theorem 3.2.1 Let m, n Z and (m, n) = 1. Then the natural map :

Z/mnZ Z/mZ Z/nZ given by : a mod mn 7 (a mod m, a mod n) yields
an isomorphism of the rings Z/mnZ and Z/mZ Z/nZ.
Proof. It is clear that is a ringhomomorphism. Furthermore is injective,
for if we assume (a) = (b) this means that a and b are equal modulo m and
n. So m and n both divide a b and since (m, n) = 1, mn divides a b, i.e
a b mod mn. Also, Z/mnZ and Z/mZ Z/nZ have the same cardinality, mn.
Since an injective map between nite sets of the same cardinality is automatically
bijective, our theorem follows.
2
Via almost the same proof we obtain the following generalisation.
Theorem 3.2.2 Let m1 , . . . , mr Z and (mi , mj ) = 1 i = j. Let m =
m1 mr . Then the map
: Z/mZ (Z/m1 Z) (Z/mr Z)
given by
: a mod m 7 a mod m1 , . . . , a mod mr
yields a ring isomorphism.
A direct consequence of Theorem 3.2.2 is as follows
3.2. CHINESE REMAINDER THEOREM
23
Corollary 3.2.3 (Chinese remainder theorem) Let notations be the same

as in Theorem 3.2.2. To any r-tuple of residue classes c1 mod m1 , . . . , cr mod mr
there exists precisely one residue class a mod m such that a ci (mod mi ) (i =
1, . . . , r).
Let us give a slightly dierent reformulation and another proof.
Theorem 3.2.4 (Chinese remainder theorem) Let m1 , . . . , mr Z and (mi , mj ) =
1 i = j. Let m = m1 mr . Then the system of equations
x cj (mod mj ),
1jr
has exactly one residue class modulo m as solution set.

Proof. Dene Mj = m/mj for j = 1, . . . , r. Then clearly (Mj , mj ) = 1, j.
Choose for each j a number nj such that Mj nj 1(mod mj ). Let
x0 =
cj nj Mj .
j=1
Using mi |Mj if i = j and ni Mi 1(mod mi ) we can observe that x0 ci (mod mi )

for any i. Clearly, any number congruent to x0 modulo m is also a solution of
our system.
Conversely, let x1 be a solution of our system. Then x1 x0 mod mj and hence
mj |(x1 x0 ) for all j. Since (mi , mj ) = 1 for all i = j, this implies m|(x1 x0 )
and hence x1 x0 mod m, as desired.
2
Notice that the latter proof of the Chinese remainder theorem gives us an algorithm to construct solutions. However, in most cases it is better to use straightforward calculation as in the following example.
Example. Solve the system of congruence equations
x 1(mod 5),
x 2(mod 6),
x 3(mod 7).
First of all notice that 5, 6 and 7 are pairwise relatively prime and so, according to
the chinese remainder theorem there is a unique residue class modulo 567 = 210
as solution. However, in the following derivation this will follow automatically.
The rst equation tells us that x = 1 + 5y for some y Z. Substitute this
into the second equation, 1 + 5y 2(mod 6). Solution of this equation yields
y 5(mod 6). Hence y must be of the form y = 5 + 6z for some z Z.
For x this implies that x = 26 + 30z. Substitute this into the third equation,
26 + 30z 3(mod 7). Solution yields z 6(mod 7). Hence z = 6 + 7u, u Z
which implies x = 206 + 210u. So the residueclass 206 mod 210 is the solution to
our problem. Notice by the way that 206 4(mod 210). If we would have been
clever we would have seen the solution x = 4 and the solution to our problem
without calculation.
24
3.3
Invertible residue classes
In this section we shall give a description of the group (Z/mZ) and a number
of properties. With the notations of Theorem 3.2.2 we know that Z/mZ
Z/m1 Z Z/mr Z. This automatically implies the following corollary.
Corollary 3.3.1 Let notations be the same as in Theorem 3.2.2. Then,
(Z/mZ) (Z/m1 Z) (Z/mr Z) .
Definition 3.3.2 Let m N. The number of natural numbers m relatively
prime with m is denoted by (m), Eulers totient function.
Notice that if m 2 then (m) is precisely the cardinality of (Z/mZ) . We recall
the following theorem which was already proved in the chapter on arithmetic
functions. However, the multiplicative property follows also directly from the
chinese remainder theorem. So we can give a separate proof here.
Theorem 3.3.3 Let m, n N.
a) If (m, n) = 1 then (mn) = (m)(n).
b) (n) = n (1 p1 ), where the product is over all primes p dividing n.
c)
d|n (d) = n.
Proof. Part a) follows directly from Corollary 3.3.1 which implies that (Z/mnZ)
(Z/mZ) (Z/nZ) . Hence |(Z/mmZ) | |(Z/mZ) | |(Z/nZ) | and thus
(mn) = (m)(n).
Part b) First note that (pk ) = pk pk1 for any prime p and any k N. It is
simply the number of integers in [1, pk ] minus the number of multiples of p in the
same interval. Then, for any n = pk11 pkr r we nd, using property (a),
(n) =
(pki i )
i=1
(pki i pki i 1 ) = n
i=1
(1
Part c) Notice that

n = #{k|1 k n}
=
#{k|1 k n, (k, n) = d}
d|n
#{l|1 l n/d, (l, n/d) = 1}
d|n
(n/d) =
d|n
d|n
(d)
1
).
pi
3.3. INVERTIBLE RESIDUE CLASSES
25
Theorem 3.3.4 (Euler) Let a Z and m N. Suppose (a, m) = 1. Then

a(m) 1(mod m).
Proof. Since (a, m) = 1, the class a mod m is an element of (Z/mZ) , a nite
abelian group. Since (m) = #(Z/mZ) , our statement follows from Lemma 13.1.1 2
When m is a prime, which we denote by p, we have (p) = p 1 and thus the
following corollary.
Corollary 3.3.5 (Fermats little theorem) Let a Z and let p be a prime.
Suppose p |a. Then
ap1 1(mod p).
Corollary 3.3.5 tells us that 2p1 1(mod p) or equivalently, 2p 2(mod p),
for any odd prime p. Almost nothing is known about the primes for which
2p 2(mod p2 ). Examples are given by p = 1093 and 3511. No other such
primes are known below 3 109 . It is not even known whether there exist
nitely or innitely many of them. And to add an even more remarkable token
of our ignorance, neither do we know if there innitely many primes p with
2p 2(mod p2 ).
Theorem 3.3.6 (Wilson,1770) If p is a prime then (p 1)! 1(mod p).
Proof. Notice that this theorem is just a special case of Lemma 13.1.12, when
we take R = (Z/pZ) .
2
Theorem 3.3.7 (Gauss) Let p be a prime. Then the group (Z/pZ) is cyclic.
Proof. Since p is prime, (Z/pZ) is the unit group of the eld Z/pZ. From
Lemma 13.1.11 our statement follows.
2
Definition 3.3.8 Let m N. An integer g such that g mod m generates the
group (Z/mZ) is called a primitive root modulo m.
As an illustration consider (Z/17Z) and the powers of 3 modulo 17,
31 3
32 9 33 10 34 13 35 5 36 15
37 11 38 16 39 14 310 8 311 7 312 4
313 12 314 2 315 6 316 1
Observe that the set {31 , . . . , 316 } equals the set {1, 2, . . . , 16} modulo 17. So 3
is a prinitive root modulo 17. Notice also that 24 16 1(mod 17). Hence
28 (1)2 1(mod 17) and 2 is not a primitive root modulo 17.
26
Not all values of m allow a primitive root. For example, the order of each elements
in (Z/24Z) is 1 or 2, whereas this group contains 8 elements. To get a good
impression of (Z/mZ) one should take the prime decomposition m = pk11 pkr r
of m and notice that according to Corollary 3.3.1
(Z/mZ) (Z/pk11 Z) (Z/pkr r Z) .
The structure for the groups (Z/pk Z) with p prime and k N is then given in
the following two theorems.
Theorem 3.3.9 Let p be an odd prime and k N. Then (Z/pk Z) is a cyclic
group.
Proof. For k = 1 the theorem is just Theorem 3.3.7. So let us assume k > 1.
Let g be a primitive root modulo p and let ord(g) be its order in (Z/pk Z) . Since
g ord(g) 1(mod pk ) we have g ord(g) 1(mod p). Moreover, g is a primitive root
modulo p and thus (p 1)|ord(g). So h = g ord(g)/(p1) has order p 1 in (Z/pk Z) .
k1
From Lemma 3.3.11(i) with r = k it follows that (1 + p)p 1(mod pk ) and the
k2
same lemma with r = k 1 implies (1 + p)p
1 + pk1 1(mod pk ). Hence
ord(1 + p) = pk1 . Lemma 13.1.6 now implies that (p + 1)h has order (p 1)pk1
and so (Z/pk Z) is cyclic.
2
Theorem 3.3.10 Let k Z3 . Any element of (Z/2k Z) can be written uniquely
in the form (1)m 5t mod 2k with m {0, 1}, 0 t < 2k2 .
Note that this theorem implies that (Z/2k Z) is isomorphic to the product of a
cyclic group of order 2 and a cyclic group of order 2k2 when k 3. Of course
(Z/4Z) and (Z/2Z) are cyclic.
Proof. From Lemma 3.3.11(ii) with r = k we nd 52
1(mod 2k ) and with
2k3
k1
k
r = k 1 we nd 5
1+2
1(mod 2 ). So, ord(5) = 2k2 . Notice that
all elements 5t , 0 t < 2k2 are distinct and 1(mod 4). Hence the remaining
elements of (Z/2k Z) are given by 5t , 0 t < 2k2 .
2
k2
Lemma 3.3.11 Let p be an odd prime and r N.

r1
i. (1 + p)p
r2
ii. 52
1 + pr (mod pr+1 ).
1 + 2r (mod 2r+1 ) for all r 2.
Proof. i. We use induction on r. For r = 1 our statement is trivial. Let r > 1

and assume we proved
r2
(1 + p)p
1 + pr1 (mod pr ).
3.4. PERIODIC DECIMAL EXPANSIONS

r2
In other words, (1 + p)p

on both sides,
pr1
(1 + p)
= 1 + Apr1 with A 1(mod p). Take the p-th power
= 1+
p ( )
p
t=1
1 + pAp
Because p is odd we have
27
(p)
2
r1
(1 + p)p
t
r1
(Apr1 )t
( )
p
+
(Apr1 )2 (mod pr+1 ).
2
0(mod p) and we are left with

1 + Apr 1 + pr (mod pr+1 )
as asserted.
ii. Use induction on r. For r = 2 our statement is trivial. Let r > 2 and assume
we proved
r3
52 1 + 2r1 (mod 2r ).
r3
In other words, 52
52
r2
= 1 + A2r1 with A odd. Take squares on both sides,

= 1 + A2r + A2 22r2 1 + 2r (mod 2r+1 ).
The latter congruence follows because A is odd and 2r 2 r + 1 if r > 2. This

concludes the induction step.
2
3.4
Periodic decimal expansions
A very nice application of the properties of residue classes is that in writing

decimal expansions of rational numbers. For example, when we write out the
decimal expansion of 1/7 we very soon nd that
1
= 0.142857142857142857142857 . . . ,
7
in short hand notation
1
= 0./14285/7.
7
Using Eulers Theorem 3.3.4 this is easy to see. Note that 106 1 is divisible by
7 and 106 1 = 7 142857. Then,
142857
1
= 6
= 142857 106 + 142857 1012 +
7
10 1
The second equality is obtained by expanding 1/(106 1) = 106 /(1 106 ) in
a geometrical series.
28
Definition 3.4.1 A decimal expansion is called periodic if it is periodic from

a certain decimal onward, and purely periodic if it is periodic starting from the
decimal point.
For any periodic decimal expansion there is a minimal period and it is clear that
the minimal period divides any other period.
Theorem 3.4.2 Let be a real number then,
i. The decimal expansion of is periodic Q.
ii. The decimal expansion of is purely periodic Q and the denominator
of is relatively prime with 10.
Suppose Q has a denominator of the form 2k 5l q, where (q, 10) = 1. Then
the minimal period length of the decimal expansion of equals the order of 10 in
(Z/qZ) .
Proof. Suppose that the decimal expansion of is periodic. By multiplication
with a suciently large power 10k we can see to it that 10k is purely periodic.
Suppose the minimal period length is l. Then there exist integers A, N such that
0 N < 10l 1 and 10k = A + N 10l + N 102l + . The latter sum equals
A + N/(10l 1), hence is rational. In case the decimal expansion is purely
periodic we can take k = 0 and we have = A + N/(10l 1). Thus it is clear
that the denominator q of is relatively prime with 10.
Suppose conversely that Q. After multiplication by a suciently high power
10k we can see to it that the denominator q of 10k is relatively prime with 10.
Then we can write 10k = A + a/q, where A, a Z and 0 a < q. Let r be the
order of 10 in (Z/qZ) and write N = a (10r 1)/q. Then 0 N < 10r 1 and
10k = A +
N
.
1
10r
After expansion into a geometrical series we see that 10k has a purely periodic
decimal expansion and that has a periodic expansion. In particular, when the
denominator of is relatively prime with 10 we can take k = 0 and the expansion
of is purely periodic in this case.
To prove the last part of the theorem we make two observations. Let again r
be the order of 10 in (Z/qZ) and let l be the minimal period of the decimal
expansion. From the rst part of the above proof it follows that q divides 10l 1.
In other words, 10l 1(mod q) and hence r|l. From the second part of the proof
it follows that r is a period of the decimal expansion, hence l|r. Thus it follows
that r = l.
2
The above theorem is stated only for numbers written in base 10. It should be
clear by now how to formulate the theorem for numbers written in any base.
3.5. EXERCISES
3.5
29
Exercises
Exercise 3.5.1 Let n N be composite and n > 4. Prove that (n 1)!

0(mod n).
Exercise 3.5.2 Prove that a natural number is modulo 9 equal to the sum of its
digits.
Exercise 3.5.3 Prove that any palindromic number of even length is divisible by
11. (A palindromic number is a number that looks the same whether you read
from right to left or from left to right).
Exercise 3.5.4 Determine the inverse of the following three residue classes:
5(mod 7), 11(mod 71), 86(mod 183).
Exercise 3.5.5 A ea jumps back and forth between the numerals on the face of
a clock. It is only able to make jumps of length 7. At a certain moment the ea
is located at the numeral I. What is the smallest number of jumps required to
reach XII?
Exercise 3.5.6 Determine all x, y, z Z such that
a) x 3(mod 4)
b) 3y 9(mod 12)
c) z 1(mod 12)
x 5(mod 21)
x 7(mod 25)
4y 5(mod 35)
z 4(mod 21)
6y 2(mod 11)
z 18(mod 35).
Exercise 3.5.7 Determine all integral multiples of 7 which, after division by

2,3,4,5 and 6 yield the remainders 1,2,3,4 and 5 respectively.
Exercise 3.5.8 Prove that there exist a million consecutive positive integers,
each of which is divisible by a square larger than 1.
Exercise 3.5.9 Notice that 3762 376(mod 103 ), 906252 90625(mod 105 ).
a) Let k N. How many solutions a Z with 1 < a < 10k does a2 a(mod 10k )
have?
b) Solve the equation in a) for k = 12.
Exercise 3.5.10 For which n N does the number nn end with 3 in its decimal
expansion?
Exercise 3.5.11 (XXI st International Mathematics Olympiad 1979). Let N, M
N be such that (N, M ) = 1 and
1 1 1
1
1
1
N
= 1 + + +
+
.
M
2 3 4
1317 1318 1319
Show that 1979 divides N . Generalise this.
30
Exercise 3.5.12 Write the explicit isomorphism between Z/10Z and Z/2Z
Z/5Z.
Exercise 3.5.13 a) Prove that 13|(42
9
b) Prove that 37/|x9 + 4 x N.
2n+1
3) n N.
Exercise 3.5.14 Suppose m = pk11 pkr r with pi distinct primes and ki 1 i.

Put
(m) = lcm[pk11 1 (p1 1), . . . , pkr r 1 (pr 1)].
Show that a(n) 1(mod n) for all a such that (a, n) = 1.
Exercise 3.5.15 Show that n13 n(mod 2730) for all n Z.
Exercise 3.5.16 For which odd primes p is (2p1 1)/p a square ?
Exercise 3.5.17 Let p be a prime such that p 1(mod 4). Prove, using Wilsons
theorem, that ((p 1)/2)!)2 1(mod p).
Exercise 3.5.18 Is the following statement true: every positive integer, written
out decimally, is either prime or can be made into a prime by changing one digit.
Exercise 3.5.19 Find a prime divisor of 1111111111111 and of 111 . . . 111 (79
ones).
Exercise 3.5.20 Prove that for all a, n N with a = 1 we have n|(an 1).
Exercise 3.5.21 Prove that (n) as n .
Exercise 3.5.22 (H.W.Lenstra jr.) Let N = 16842752.
a) Find an even n N such that (n) = N .
b) Prove that there is no odd number n N such that (n) = N .
c) Let 1 k 31. Construct an odd m N such that (m) = 2k .
Exercise 3.5.23 Prove that for all n 2,
n1
m=1
(m,n)=1
1
m = n(n).
2
Exercise 3.5.24 Prove that

{
0 if n N is even
(1)n/d (d) =
n if n N is odd
d|n
3.5. EXERCISES
31
Exercise 3.5.25 a) Determine all solutions to (n) = 24.

b) Show that there are no solutions to (n) = 14.
Exercise 3.5.26 Given a N characterise all n N such that (n)/n = (a)/a.
Exercise 3.5.27 Determine all n N such that (n)|n.
Exercise 3.5.28 Compute the orders of 5(mod 7), 3(mod 46), 26(mod 17) in
the multiplicative groups mod m.
Exercise 3.5.29 Determine all primes such that thet period of the decimal expansion of 1/p is
a) at most 6
b) precisely 7
c) precisely 8.
Exercise 3.5.30 Let p be an odd prime and q a prime divisor of 2p 1. Prove
that q = 2mp + 1 for some m N.
Exercise 3.5.31 Let n N, a Z and q an odd prime divisor of a2 + 1.
a) Prove that q = m 2n+1 + 1 for some m N.
b) Prove that for any n N there exist innitely many primes of the form p
1(mod 2n ).
n
Exercise 3.5.32 Let q be an odd prime and n N. Show that any prime divisor
p of nq1 + + n + 1 satises either p 1(mod q) or p = q. Using this fact
show that there exist innitely many primes which are 1 modulo q.
Exercise 3.5.33 Determine all primitive roots modulo 11. Determine all primitive roots modulo 121.
Exercise 3.5.34 a) Determine all primitive roots modulo 13, 14 and 15 respectively.
b) Solve the following equations: x5 7(mod 13), x5 11(mod 14), x5
2(mod 15).
Exercise 3.5.35 Suppose that (Z/mZ) is cyclic. Prove that m = 2, 4, pk or 2pk ,
where p is an odd prime and k N. (Hint: use Exercise 3.5.14)
Exercise 3.5.36 Suppose that p and q = 2p + 1 are odd primes. Prove that
p1
(1) 2 2 is a primitive root modulo q.
Exercise 3.5.37 Show that 22 1 is divisible by at least t distinct primes.
t
32
Exercise 3.5.38 In the decimal expansion of 1/5681 the 99-th digit after the
decimal point is a 0. Prove this.
Exercise 3.5.39 What is the 840-th digit after the decimal point in the decimal
expansion of 1/30073?
Exercise 3.5.40 What is the 165-th digit after the decimal point in the decimal
expansion of 1/161?
Exercise 3.5.41 (Lehmers prime test). Let a, n N be such that
an1 1(mod n) and
n1
p
1(mod n) primes p|(n 1).
Prove that n is prime.

Exercise 3.5.42 Use Lehmers test to prove that 3 28 + 1 is prime. (Hint:use
a = 11). Notice that Lehmers test is very suitable for numbers such as h 2k + 1
when h is small.
Chapter 4
Primality and factorisation
4.1
Prime tests and compositeness tests
It is well known that one can nd the primefactorisation of a numbers N by the

naive method of trying all numbers below N as divisor. It is also known that
it is extremely time-consuming to factor any number with more than 10 digits in
this way. As an example, try to factor N = 204906161249 in this way by hand!
When the number of digits of N exceeds 20 digits this naive method consumes
hours of computer time, even on a fast PC. During the last 20 years powerful
new methods have been developed to factorise large numbers more eciently. A
recent success has been the factorisation of the Fermat number 2512 + 1, which
is now, and probably for a long time to be, the largest Fermat number of which
we know the complete prime factorisation. Nowadays (around 2012) numbers
of about 200 digits can be factored fairly routinely. Before one tries to factor a
number however, it makes sense to check rst whether it is prime. Such primality
tests have been developed as well and using routine methods on a supercomputer
one can test primality of numbers up to 1000 digits. In this section we describe
a few simple tests which are in practice very ecient and can be programmed
easily on a PC.
Consider our number N again. We would like to know
whether it is prime or
not. Again, the naive method of trying all numbers below N as divisor becomes
impossible if N contains more than 20 digits. The new idea is to use the following
contrapositive to Fermats little theorem.
Theorem 4.1.1 Let N N and let a Z be such that N does not divide a.
Then,
aN 1 1(mod N ) = N is composite.
So, given N , pick any a not divisible by N and perform the above test. However,
before trying to apply this test, it is sensible to check whether (a, N ) = 1. If not,
N is of course composite and we have also found a divisor. In that case there is
33
34
CHAPTER 4. PRIMALITY AND FACTORISATION
no need to test N any further. As an example of the evaluation of aN 1 (mod N ),

take a = 2 and N = 204906161249. The computation of 2N 1 mod N may seem
formidable at rst sight, but with a little thought it can be done very eciently.
To do this write N 1 in base 2,
N 1 = 10111110110101010111110010001100000.
We perform a number of squaring operations and multiplications by 2 in such a
way, that we try to form the binary expansion of N 1 in the exponent of 2 as
we go along,
21
210
2101
21011
2N 1
2(mod N )
22 4(mod N )
2 42 32(mod N )
2 322 2048(mod N )
201135347146(mod N ).
The number of steps required by this method is precisely the number of binary
digits of N 1 which isproportional to log N . For large N this is extremely
small compared to the N steps required by the naive method. By the way,
we see that 2N 1 1(mod N ), hence N is composite (it will turn out that
N = 369181 555029). So Theorem 4.1.1 can be considered as a compositeness
test. What to do if we had found 2N 1 1(mod N ) instead? We cannot conclude
that N is prime. We have for example 2560 1(mod 561), whereas 561 = 31117.
But we can always choose other a and repeat the test. If aN 1 1(mod N ) for
several a, we still cannot conclude that N is prime. It is becoming more likely,
but not 100% certain. In fact there exist N such that aN 1 1(mod N ) for
all a with (a, N ) = 1. These are the so-called Carmichael numbers. Examples
are 561, 1729, 294409, . . .. There exist 2163 Carmichael numbers below 25 109 .
It was an exciting surprise when Granville, Pomerance and Red Alford showed
around 1991 that there exist innitely many of them.
A criterion which gives better chances (but not certainty) in recognising primes
is based on the following renement of Fermats little theorem.
Theorem 4.1.2 Let p be an odd number and a Z such that p |a. Write
p 1 = 2k m with m odd and k 0. Suppose p is prime. Then,
either am 1(mod p)
or i such that a2 m 1(mod p) and 0 i k 1.
i
Proof. Since p is prime we know that

k m
ap1 a2
1(mod p).
4.1. PRIME TESTS AND COMPOSITENESS TESTS
35
Suppose that am 1(mod p). Let r be the smallest non-negative integer such
r
r1
that a2 m 1(mod p). Notice that 1 r k. Then a2 m 1(mod p).
r1
r1
By the minimality of r we cannot have a2 m 1(mod p). Hence a2 m
1(mod p) and since 0 r 1 k 1 our assertion follows.
2
The contrapositive statement can be formulated as follows.
Theorem 4.1.3 (Rabin test) Let N N be odd and a Z such that N /
|a.
Write N 1 = 2k m with k 0 and m odd. If
am 1(mod N )
and
0ik1 : a2 m 1(mod N )
i
(4.1)
then N is composite.
If a satises property (4.1) we shall call a a witness to the compositeness of
N . Unlike the converse to Fermats little theorem the Rabin test allows no
Carmichael-like numbers N . This is garantueed by the following theorem
Theorem 4.1.4 Let N N be odd and composite. Among the integers between
1 and N at least 75% is a witness to the compositeness of N .
Proof. It suces to prove that at least 75% of the numbers in (Z/N Z) is a
witness. We shall do this for all N = 9. For N = 9 the theorem is directly veried
by hand. Let S be the union of the solution sets of the equations xm 1(mod N )
i
and x2 m 1(mod N )(i = 0, . . . , k 1) respectively. It suces to show that
|S| is at most (N )/4.
j
Let j be the largest number with 0 j k 1 such that x2 m 1(mod N ) has
a solution. Such a j exists since we have trivially (1)m 1(mod N ). Notice
j
that S is contained in the set of solutions of x2 m 1(mod N ). Now apply
Lemma 4.1.5 to see that there are at most (N )/4 such solutions.
2
Lemma 4.1.5 Let N be an odd composite positive integer, not equal to 9. Let
M |(N 1)/2. Suppose that xM 1(mod N ) has a solution x0 . Then the
number of solutions to xM 1(mod N ) in x (Z/N Z) is less than or equal
to (N )/4.
Proof. Let N = pk11 pkr r be the prime factorisation of N . Notice that the
number of solutions to xM 1(mod N ) equals the numbers of solutions to xM
1(mod N ), a bijection being given by x 7 xx0 (mod N ). The total number of
solutions is equal to
2 (M, pki i 1 (pi 1)) = 2 (M, pi 1).

i
The equality follows from the fact that, since M |(N 1), prime factors of N
cannot divide M . For every prime p dividing N we know that xM 1(mod p)
36
has a solution, hence ord(x) divides 2M but not M . This implies (M, p 1) =
ord(x)/2 (p 1)/2.
Suppose that N has at least three distinct prime factors. Then,
pi 1
1
(N )
2 (M, pi 1) 2
(pi 1)
.
2
4 i
4
i
i
Suppose that N has precisely two distinct prime factors, p and q say. First
suppose that N = pq. There exist e, f N such that p 1 = 2e(M, p 1) and
q 1 = 2f (M, q 1). Suppose e = f = 1. From e = 1 follows that (p 1)/2
divides M which in its turn divides (N 1)/2 = (pq 1)/2. Hence (p 1)/2
divides (q 1)/2. Similarly it follows from f = 1 that (q 1)/2 divides (p 1)/2.
Hence p = q, which is impossible. So we must assume ef 2. But then,
2(M, p 1)(M, q 1) = 2
(p 1)(q 1)
(p 1)(q 1)
(N )
=
.
4ef
4
4
Now suppose that N has two primes p, q and is divisible by p2 , say. Then,
2(M, p 1)(M, q 1)
2(p 1)(q 1)
(N )
(N )
.
4
2p
4
Finally suppose that N is a prime power, pk , k 2. Then, since pk1 4,

2(M, p 1) (p 1)
(N )
(N )
.
k1
p
4
2
In practice we choose a arbitrarily and apply Rabins test. Suppose that N

is composite. Let us assume that the witnesses to the compositeness of a are
distributed more or less randomly, which does not seem to be unreasonable.
Then, according to Theorem 4.1.3 the chance that N will not be recognised as
composite is less than 1/4. This means that after a 100, say, of such trials the
chance that N is not recognised as composite is less than (1/4)100 , very small
indeed. So, alternatively, if a number N is not recognised as composite after a
100 trials, the likelyhood that it is prime is practically 1!
Under the assumption of a certain conjecture Rabins test can be made into a
fulledged primality test.
Theorem 4.1.6 Let N be a composite number. If the Generalised Riemann Hypothesis (GRH) is true, then there is witness to the compositeness of N which is
smaller than 2(log N )2 .
In other words, under the assumption of GRH, if N has no witness below 2(log N )2
then it is prime. This is known as the Rabin-Miller test and its run-time has the
pleasant property of being polynomial in the length of the input (i.e. polynomial
in log N ). Unfortunately, there is not much hope of proving the GRH yet.
A nice primality test is the following.
4.1. PRIME TESTS AND COMPOSITENESS TESTS
37
Theorem 4.1.7 (Lehmer) Let N N. Suppose that we have a number a such

that
aN 1 1(mod N )
a(N 1)/q 1(mod N ) for every prime q dividing N 1.
Then N is prime.
Proof. Let k be the order of a(mod N ). From the rst property it follows that k
divides N 1. Suppose that k equals N 1. Then we conclude that (N ) N 1
and this is only possible if N is prime. So we are done when k = N 1.
Now suppose k < N 1. Then there exists a prime divisor p of N 1 such
that a(N 1)/p 1(mod N ). But this contradicts our second assumption on a. So
k < N 1 cannot occur.
2
For numbers of a very special form there may exist very powerful primality tests
tailored to the special circumstances. We mention here the Fermat numbers
n
22 + 1 and the Mersenne numbers 2n 1. The following theorem can be seen
as a very special case of Lehmers test where N 1 has only prime factors 2. It
turns out that a = 3 has the desired properties.
Theorem 4.1.8 (Pepin, 1877) Let n N and Fn = 22 + 1. Then,
n
Fn is prime 3
Fn 1
2
1(mod Fn ).
Theorem 4.1.9 (Lucas, 1876) Let Mn = 2n 1. Consider the sequence S1 , S2 , S3 , . . .

2
given by S1 = 4 and Sk = Sk1
2 for all k 2. Then,
Mn is prime Sn1 0(mod Mn ).
The proofs of these theorems are given in the chapter on quadratic reciprocity.
It turns out that F1 , F2 , F3 , F4 are prime. This led Fermat to believe that all
numbers Fn are prime. Ironically it turns out that Fn is composite for n =
5, 6, . . . , 22 and many other n. It is not even known whether Fn is prime for any
n > 4. Here is a list of factorisations of Fn for 5 n 9,
F5 = 641 6700417 (Euler, 1732)
F6 = 274177 67280421310721 (Landry, Le Lasseur, 1880)
F7 = 59649589127497217
5704689200685129054721 (Morrison, Brillhart, 1974)
F8 = 1238926361552897
(9346163971535797776916355819960689658405123754163
8188580280321) (Brent, Pollard, 1980)
38

F9 = 2424833
7455602825647884208337395736200454918783366342657
(7416400626753080152478714190193747405994078109751
9023905821316144415759504705008092818711693940737)
(Lenstra, Manasse, 1990)
In 1644 The French monk Marin Mersenne stated that 2n 1 is prime for the
values
n = 2, 3, 5, 7, 13, 17, 19, 31, 67, 127, 257
and composite for all other n < 257. It was clear that Mersenne had not tested
all these numbers. It was only in 1750 when Euler veried that 231 1 is prime.
It also turned out that Mersenne was wrong about n = 67, 257 and that he had
forgotten to add n = 61, 89, 107 to his list. The number 2127 1 was proven to be
prime in 1876 by Lucas and until 1952 this remained the largest known prime. For
further details and more on prime numbers see the Web page primes.utm.edu.
For the following values of n the number Mn is now (2013) known to be prime:
n=
2, 3, 5, 7, 13, 17, 19, 31, 61, 89, 107, 127, 521, 607, 1279, 2203,
2281, 3217, 4253, 4423, 9689, 9941, 11213, 19937, 21701,
23209, 44497, 86243, 110503, 132049, 216091, 756839, 859433,
1257787, 1398269, 2976221, 3021377, 6972593, 13466917,
20996011, 24036583, 25964951, 30402457, 32582657,
37156667, 42643801, 43112609, 57885161
At the moment (2013) 257885161 1 is the largest known prime.

For the latest news on search activities see: primes.utm.edu/mersenne/ or
www.mersenne.org.
4.2
A polynomial time primality test
In recent years, starting from the 1980s, several powerful primality tests have
been invented. We mention the test of Adleman,Rumely,Lenstra and Cohen,
which uses Gauss sums (see the chapter on Gauss sums) and tests which use the
addition structure on elliptic curves and abelian varieties. These methods are
still used in practice.
Despite all these ingeneous developments it still was not clear whether there exists a primality test whose runtime is polynomial in the number of digits of the
number to be tested. This changed in July 2002. An Indian computer scientist,
M.Agrawal and two of his students, N.Kayal and N.Saxena, had discovered a
polynomial time primality test. This was a historical breakthrough in the theory of factorisation and primality proving. Another remarkable feature of their
4.3. FACTORISATION METHODS
39
discovery was its elementary nature. The original ingredients were some modular arithmetic with polynomials and a deep theorem in analytic number theory.
However, through the eorts of H.W.Lenstra jr, the analytic number theory part
has been replaced by a property of prime numbers which will actually be proved
in these course notes.
To show that the algorithm is quite simple we give it here is pseudo-code.
Input: integer n > 1
1. If n = ab for a N and b > 1:
output Composite
2. Find the smallest r N such that ordr (n) > 4(log n)2 .
3. If 1 < gcd(a, n) < n for some a r:
output Composite
4. If n r, output Prime
5. For a = 1 to [2 (r) log n] do

If (X + a)n = X n + a(mod X r 1, n):
output Composite
6. output Prime
Here is a proof for the correctness of the algorithm. In Step 1. it is clear that if
n = ab with b > 1, then n is composite. Furthermore, in Step 3 it is clear that
if gcd(a, n) is not 1 or n, the number n is composite. For Step 5 we remark that
if n is prime, then (X + a)n X n + a(mod n) for all integers a. So, a fortiori,
(X + a)n X n + a(mod X r 1, n) for all r and a. If this condition is violated
for some a, r, then of course n cannot be prime.
The main point is now to show that
if (X + a)n X n + a(mod X r 1, n) for our
chosen r and all a between 1 and [2 (r) log n], then n is prime.
4.3
Factorisation methods
The test of Rabin, described in the previous section, enables one to decide compositeness of a number N without knowing anything about the prime decomposition of N . In this section we make a few remarks about the problem of factoring
numbers. First of all, factoring a number is much harder than proving its compositeness. Another feature of factorisation algorithms is that they mostly give
a positive chance of success, but not certainty. Factorisation methods with 100%
of success are the naive method and the algorithm of Sherman-Lehman. However
they are both very slow with runtimes O(N 1/2 ) and O(N 1/3 ) respectively. One
therefore prefers the probabilistic methods with the losophy that if you happen
40
to nd a factorisation of a large number you dont complain about how it was

found. As an example of such a factorisation method we sketch the Pollard rho
algorithm. It works very well on a PC for numbers up to 25 digits.
The Pollard rho method. Choose b Z and b = 0, 2. Compute the numbers
x0 , x1 , x2 , . . . by
x0 = 1 and xk+1 x2k + b(mod N ) for k 0
(4.2)
For each k check whether

N > gcd(xk xl , N ) > 1
(4.3)
for some l < k. If condition (4.3) is fullled we have actually found a factor of N .
Let p be a prime divisor
of N . We assert that the probability to nd p using this
algorithm with k < 2pis larger than 1/2. Since a composite number N always
has a prime divisor < N , the expected run-time of our algorithm is therefore
O(N 1/4 ).
Here is a heuristic argument which supports our assertion. Consider the sequence
x0 , x1 , x2 , . . . modulo p. Then practical experience suggests that if b = 0, 2, this
sequence behaves randomly modulo p. People are unable to prove this but it
seems like a good principle. Let q N. Being a random sequence, the probability
that the elements x0 , x1 , . . . , xq mod p are all distinct is
1
2
q
(1 )(1 ) (1 ).
p
p
p
Suppose that q >
2p. Then
1
2
q
r
log(1 )(1 ) (1 ) =
log(1 )
p
p
p
p
r=1
r
r=1
1 q(q + 1)
< 1
2
p
and the probability

that x0 , . . . , xq mod p are distinct is less than e1 < 1/2. So,
when q > 2p the probality that two elements in x0 , . . . , xq mod p are equal
is larger than1/2. In other words, if we are moderately lucky we will nd
0 l < k 2p such that xk xl (mod p). Note by the way, that the above
argument is an example of the so-called birthday paradox.
If we carry out the algorithm as described above, we would have to store the
elements xi of our sequence to verify (4.3). Moreover, testing all l < k with
4.4. THE QUADRATIC SIEVE
41
respect to (N, xk xl ) would considerably slow down the algorithm. In practice

there are two ways to overcome this problem.
The rst is to compute
xi and x2i simultaneously at every step. Suppose there
exist l < k < 2p satisfying xk xl (mod p). By its very construction, the
sequence {xi mod p}
i=0 will be periodic from the index l onward with period
k l. In particular, choose M N such that k > M (k l) l. Then we have
by the periodicity, x2M (kl) xM (kl) (mod p) and
we nd that p|(x2i xi , N ) for
i = M (k l). Notice that i = M (k l) < k < 2p.
A second option is to compute the sequence of xi (mod N ) and whenever we hit
upon i being a power of 2 we save it as the number A. At every iteration we
check whether 1 < gcd(xi A) < N . Suppose that from the index l onward the
sequence xi (mod p) becomes periodic with period k l. Choose r minimal such
that 2r max l, k l, then there exists i with i < 2r+1 such that xi x2r
xi A 0(mod p).
It should also be noted that instead of xk+1 = x2k + b we could have chosen any
other recurrence which has a chance of producing random sequences modulo N .
Our choice is simply the simplest we could think of.
In many large factorisation programs one factors out small prime factors by native
trial division for primes up to 109 , say. As a second step one often uses the Pollard
rho method to nd moderately small prime factors up to, say 1014 .
The factorisation algorithms which are the most powerful at the moment (2006)
are the quadratic sieve of Pomerance (1984, to be discussed in the next section),
the elliptic curve method of Lenstra (1984), the Number Field Sieve (1990) by
the Lenstras, Manasse and Pollard and variations on these methods. Typically,
these methods do not garantuee 100% success in factoring a number, and their
run-time analysis is again based on probability arguments. But most of the time
they are very successful.
4.4
The quadratic sieve
From time to time even Fermat was forced to factor large numbers during his
calculations. Of course he also stumbled upon the near impossibility of fcatoring
large numbers. However, Fermat did make a few observations which enabled his
to factor certain large numbers quickly. One such example is N = 8051. Fermat
noticed that this is the dierence of two squares, 8051 = 902 72 and he got
8051 = (90 7)(90 + 7) = 83 97.
This idea can
be formalised as follows, which we call2 the method 2of Fermat.
Choose r = [ N ] and test if one of the numbers (r + 1) N, (r + 2) N, (r +
3)2 N, . . . is a square. If this is the case, say (r + k)2 N = m2 then we have
N = (r + k)2 m2 = (r + k m)(r + k + m) and thus a factorisation. The
dierence between the factors is 2m. This means that if N = ab with a < b, then
42
m = (b a)/2. Hence
k = b m r = b (b a)/2 [ N ] < (a + b)/2 ab + 1 = 1 + ( b a)2 /2.

From this we see that Fermats method cannot only work well if the dierence
between a and b is small. If one is lucky this occurs and in this way Fermat
managed to factor numbers successfully. For example, the fcatorisation
2027651281 = 46061 44021
was found in only 12 steps. Unfortunately, for large dierences Fermats method
fails completely. It may even be worse than naive factorisation.
In the history of number factoring there are many variations on Fermats method.
They all come down to the construction, in one way or another, of two integers
x and y such that x2 y 2 (mod N ). From the fact that
N = gcd(N, x2 y 2 ) = gcd(N, x y)gcd(N, x + y)
follows a possible non-trivial factorisation. The best method in this respect in
C.Pomerances quadratic sieve (1984).
This method starts in the same way as
Fermats method. We choose r = [ N ] + 1 and consider the numbers
(r + k)2 N
k = 0, 1, 2, 3, . . .
We cannot assume that N is the product of two numbers of comparable size, so

we cannot expect to into a square very quickly. The idea however is that for
small k the numbers (r + k)2 N are small relative to N . More precisely,
(r + k)2 N < ( N + k)2 N = 2k N + k 2 .
For small k the order of magnitude is k N . The philosophy is that smaller

numbers have a better chance of containing small prime factors.
We choose a bound B and search for numbers (r + k)2 N all of whose prime
factors are B. To improve our chances we also
allow negative values of2k. Here
is an example. Take N = 123889. Then r = [ N ] + 1 = 352 and (r + k) N =
15 + 704k + k 2 . We choose B = 13 and try all 20 k 20. We nd the
following values of k for which k 2 + 704k + 15 consists of prime factors 13,
k
k 2 + 704k + 15
19
23 53 13
24 36
17
9
25 3 5 13
35
0
4
1
2 32 5
27 3 13
7
15
24 33 13
4.4. THE QUADRATIC SIEVE
43
The idea is now to choose factorisation on the right hand side such their product is
a square. For example the products corresponding to k = 19, 9, 15 multiplied
yield a square. Explicitly,
(r 19)2 23 53 13(mod N )
(r 9)2 25 3 5 13(mod N )
(r + 15)2 24 33 52 (mod N )
Multiplication of these congruences yields
(r 19)2 (r 9)2 (r + 15)2 (26 32 53 13)2 (mod N )
We have two dierent squares, equal modulo N . Let us see if this gives a factor
of N ,
gcd(N, (r 19)(r 9)(r + 15) 26 32 53 13) = gcd(123889, 40982373) = 541
We were lucky and found N = 541 229.
From this example the principle of the method can be deduced. We rst nd
suciently many k so that (r + k)2 N consists of prime factors B. By
sucient we mean: at least two more values than the number of primes B,
preferably more. Linear algebra over Z/2Z tells us that we can choose from these
numbers a set whose product is a square. By taking this product modulo N we
nd a relation of the form X 2 Y 2 (mod N ), from which we hope to deduce a
factorisation. The relationship with linear algebra can be seen from our example.
The factorisation table we gave before can be depicted schematically as follows.
k 1 2
19 1 1
17 1 0
9
1 1
0
0 0
0 0
1
7
0 1
15
0 0
3
0
0
1
1
0
1
1
5
1
0
1
1
1
0
0
7
0
0
0
0
0
0
0
11
0
0
0
0
0
0
0
13
1
0
1
0
0
1
0
Notice that 7 and 11 do not occur in our factorisations. Finding factorsiations

whose product is a square comes down to nding rows in our table whose sum
is the zero-vector modulo 2. In other words, we must solve a system of linear
equations over Z/2Z.
It has not been explained yet why this method is fast. A careful analysis, using
deep heuristics from analytic number theory, shows that the expected runtime of
the algorithm is L(N )c where
)
(
log(N ) log log(N )
L(N ) = exp
44
and c = 2. The reader can verify that limN L(N )/N = 0 for every > 0. So
the runtime is not exponential, we call this a sub-exponential algorithm.
One of the bottle-necks of the algorithm is that for each (r + k)2 N it must be
decided if it consists of prime factors B. It was Pomerances idea to replace this
by a sieving process, hence the name quadratic sieve. We work with a sequence
of number ak which we initially choose to be (r + k)2 N . For every prime
p B we do the following. Solve (r + x)2 N (mod p) and let x1 , x2 be two
solutions with x1 x2 (mod p). For every k with with k x1 , x2 (mod p) we
replace ak by ak /p. Having done all this, we pick those k fro which ak = 1. For
these k the number (r + k)2 N consists only of primes B. Strictly speaking
we must also sieve for higher powers of p, but we leave this aside for the sake of
simplicity. The saving in runtime is tremendous, the runtime is now L(N ) instead
of L(N )2 . As a rule of thumb we choose B < L(N )b and |k| < L(N )a for suitable
0.1 < a, b < 1, depending on the implementation. A big advantage of the sieving
technique is that it can be distributed easily over dierent machines. To give an
idea, A.K.Lenstras factorisation of RSA-129 (129 decimals) was distributed over
1600 computers, mostly workstations and PCs. The number of primes for which
there was sieved was 524339, about half a million. The lienar algebra part was
the solution of a system of linear equations modulo 2 in half a million variables.
This work must be done on a central computer.
4.5
Cryptosystems, zero-knowledge proofs
An important recent application of large primes is the cryptosystem of Rivest,

Shamir and Adleman (RSA-cryptosystem). This system consists of a public encryption key, by which anyone can encrypt messages or other information, and
a secret key which one uses to decrypt these messages again. The important
principle of the RSA-system is that one cannot deduce the secret key from the
public key. Let us give a short description and explanation.
Choose two prime numbers p, q and let N = pq. Let = (N ) = (p 1)(q 1).
Destroy the prime numbers. Notice that
a 1(mod N )
a with (a, N ) = 1.
Choose a k, l N such that kl 1(mod ).

The public key consists of the numbers k and N , which may be advertised anywhere. The secret key consists of the numbers l and N , of which l is known only
to the owner. Write kl = 1 + m. The encryption goes as follows. Transform the
message into blocks of digits and consider them as numbers, of which we assume
that they are smaller than N . Let T be such a block. The encoded form of T
will be C determined by C T k (mod N ). The decryption simply consists of
determining C l mod N , because C l T kl T m+1 T (mod N ).
4.5. CRYPTOSYSTEMS, ZERO-KNOWLEDGE PROOFS
45
Suppose we want to deduce the secret key l from the public key k. This could be
done as follows. Factorise N = pq, determine = (p 1)(q 1) and determine
l by solving kl 1(mod ). Notice that in order to determine we need the
factorisation of N . No other methods are known. If however the secret primes p, q
contain about 100 digits, factorisation of N with the known methods is practically
impossible within a human or even universal lifetime. So the success of the RSAcryptosystem is based on the apparent inability of mathematicians to factor large
numbers.
Of course the above idea can be applied in the opposite direction. We then have
the possibility of electronic signatures. Let us sketch a very simplied application. An individual, say Peter, makes or buys his own secret/public key pair.
The public key is known to, say, a savings bank and the private key is securely
guarded by Peter. Suppose Peter wants to transfer $10,000.- from the bank to a
furniture shop electronically. He might write an electronic message saying My
name is Peter, please transfer $10,000.- to the account of furniture shop so and
so. Unfortunately anyone could do this in Peters name with undesirable consequences for Peter. The solution is that Peter encrypts his message with his
private code and appends it to the message written in plain text. The bank then
decodes the scrambled message with the public key and observes that the result
corresponds with Peters plain text. Since no one else but Peter could have encrypted the message succesfully, the bank is convinced of the validity of Peters
message and carries out his order.
Other applications of modular arithmetic and the near impossibility to factor
large numbers are protocols by which one can prove that one possesses certain
information, a password for example, without revealing that information. Such
protocols are known as zero-knowledge proofs. . One such procedure, known as
identity proof, has been devised by Goldwasser,Micali and Racko in 1985.
Suppose Vincent calls Vera over the telephone and Vincent wants to convince Vera
that it is really him, which Vera must then verify. Of course Vincent can mention
a password, only known to him and Vera, to identify himself. However, there is
always the possibility of eavesdroppers. Another problem might be that Vincent
considers Vera too sloppy to conde his password to. Both these problems are
solved by using a zero-knowledge proof.
Just as above we let N be the product of two very large primes p and q. The secret
password of Vincent, only known to him, is a number a between 1 and N . The
name by which Vincent is known publicly is A, determined by A a2 (mod N )
and 1 < A < N . The identication protocol runs as follows. Vincent takes a
random number x and sends the square X x2 (mod N ) to Vera. Then Vera can
ask either of two questions,
i. send x
ii. send ax
46
If Vincent is really Vincent he can of course do that and Vera can check that
x2 X(mod N ) in case i) and that (ax)2 AX(mod N ) in case ii).
What if an impostor tries to pass himself as Vincent? Before sending anything
he may try to guess Veras question. If he guesses i) then he can take any x send
X mod N and answer Veras question with x. If he guesses ii) he can take any
x, send A1 X mod N and answer Veras question with x. In any case the chance
for the imposter to make the right guess is 1/2. However, if this question and
answer game is repeated a hundred times, say, the chance that the impostor is
not exposed as a fraud is (1/2)100 . This is small enough for Vera to be convinced
of Vincents identity if all questions are answered correctly. Notice that in the
process the value of a has not been revealed by Vincent.
Another possibility for the imposter is to infer Vincents password a from A. He
would have to solve x2 A(mod N ) and the only known way to do this is to
solve the equivalent system x2 A(mod p), x2 A(mod q). Taking square
roots modulo a prime is doable in practice (see chapter on quadratic reciprocity)
so it seems we are done. Unfortunately we need the factorisation of N again and
this turns out to be the bottleneck in solving x2 A(mod N ). Again the safety
of Vincents password relies on our inability to factor very large numbers.
Chapter 5
Quadratic reciprocity
5.1
The Legendre symbol
In this chapter we shall consider quadratic equations in Z/mZ and study an

important criterion for the solubility of x2 a mod p, where p is an odd prime
(quadratic reciprocity).
Definition 5.1.1 Let p be an odd prime and a Z not divisible by p. Then a is
called a quadratic residue mod p if x2 a(mod p) has a solution and a quadratic
nonresidue modulo p if x2 a(mod p) has no solution.
Example. The quadratic residues modulo 13 read: 1,4,9,3,12,10 and the quadratic
nonresidues : 2,5,6,7,8,11.
Definition 5.1.2 Let p be an odd prime. The Legendre symbol is dened by
( ) 1
if a is quadratic residue mod p
a
1 if a is quadratic nonresidue mod p
=
p
0
if p|a
Theorem 5.1.3 Let p be an odd prime and a, b Z. Then
a) There are exactly
residues mod p.
p1
2
quadratic residues mod p and
b) (Euler)
( )
p1
a
a 2 (mod p).
p
c)
( )( ) ( )
a
b
ab
=
.
p
p
p
47
p1
2
quadratic non-
48
CHAPTER 5. QUADRATIC RECIPROCITY

d)
( )
1
= 1,
p
1
p
{
=
1
if p 1(mod 4)
1 if p 1(mod 4)
( )2
Proof. Part a). Consider the residue classes 12 , 22 , . . . , p1
mod p. Since a2
2
(a)2 (mod p) these are all quadratic residues modulo p. They are also distinct,
from a2 b2 (mod p) would follow a b(mod p) and when 1 a, b p1
2
this implies a = b. So there are exactly p1
quadratic
residues
modulo
p.
The
2
remaining p 1 p1
= p1
residu classes are of course quadratic nonresidues .
2
2
Part b) Clear if a 0(mod p). So assume a 0(mod p). Since (a(p1)/2 )2
ap1 1(mod p) by Fermats little theorem we see that a(p1)/2 1(mod p).
Suppose that a is a quadratic residue , i.e. x such that x2 a(mod p). Then
1 xp1 (x2 )(p1)/2 a(p1)/2 (mod p), which proves half of our assertion.
Since we work in the eld Z/pZ, the equation x(p1)/2 1(mod p) has at most
p1
solutions. We know these solutions to be the p1
quadratic residues . Hence
2
2
(p1)/2
a
1(mod p) for any quadratic nonresidue a mod p.
Part c)
( )( )
( )
p1 p1
p1
a
b
ab
a 2 b 2 (ab) 2
(mod p).
p
p
p
Because
(
) ( ) Legendre
( ) symbols can only be 0, 1 and p 3, the strict equality
a
b
= ab
follows.
p
p
p
( )
( )
Part d) Of course p1 = 1 is trivial. From part b) follows that 1

p
(1)
5.2
p1
2
(mod p). Since p 3 strict equality follows.
Quadratic reciprocity
One might wonder for which prime numbers the numbers 3 and 5 ,say, are
quadratic residue . Euler, Legendre and Gauss have occupied themselves with
this question. For example it turns out that
( )
( )
3
3
= 1 if p 1(mod 12)
= 1 if p 5(mod 12)
p
p
and
( )
5
= 1 if p 1(mod 5)
p
( )
5
= 1 if p 2(mod 5).
p
Starting from such observations Euler conjectured the quadratic reciprocity law
(see Theorem 5.2.6). Legendre gave an incomplete proof of it and later Gauss
managed to give several complete proofs. In this chapter we give a proof which
is basically a version given by Eisenstein. In the chapter on Gauss sums we shall
give another proof.
5.2. QUADRATIC RECIPROCITY
49
Definition 5.2.1 The residue classes 1, 2, . . . , p1

mod p are called positive, the
2
p1
residue classes 1, 2, . . . , 2 mod p are called negative.
Theorem 5.2.2 (Gauss Lemma) Let p be an odd prime and a Z not divisible by p. Then
( )
a
= (1)
p
where is the number of negative residue classes from a, 2a, . . . , p1
a mod p.
2
Proof. Consider the p 1 residue classes a, 2a, . . . , p1
a mod p. They
2
are non-zero and mutually distinct. So, the sets {a, 2a, . . . , p1
a} and
2
p1
p1
{1, 2, . . . , 2 } are equal modulo p. Of each pair 1, 2, . . . , 2 mod p
exactly one element occurs in the sequence a, 2a, . . . , p1
a mod p. Thus we nd
2
that
p1
p1
a 2a
a (1) 1 2
(mod p)
2
2
and hence
)
(
)
(
p1
p1
p1
!a 2 (1)
!(mod p).
2
2
p1
2
After division by ( p1
)! we obtain
(1) mod p and after using Theo2
( ) a
rem 5.1.3(b) we conclude that ap = (1) .
2
Theorem 5.2.3 Let p be an odd prime. Then

( ) {
2
1
if p 1(mod 8)
=
1 if p 3(mod 8)
p
Proof. We apply Gauss lemma. To do so we must count , the number of
negative residue classes among 2, 4, . . . , p 1 mod p. So,
p+1
n p 1}
2
p+1
p1
n
}
= #{n|
4
2
= #{n even|
Replace n by
p+1
2
n to obtain
= #{n | 1 n
[
]
p+1
=
4
p+1
}
4
This implies that is even if p 1(mod 8) and is odd if p 3(mod 8).

Gauss lemma now yields our assertion.
2
50
Remark 5.2.4 Notice that
( )
p2 1
2
= (1) 8 .
p
Another consequence of Gauss lemma is the following lemma which will be

needed in the proof of the quadratic reciprocity law.
Lemma 5.2.5 Let p be an odd prime and a Z odd and not divisible by p.
Dene
p1
]
2 [
as
S(a, p) =
.
p
s=1
( )
a
= (1)S(a,p) .
p
( )
Proof. According to Gauss lemma we have ap = (1) where is the number
Then
a mod p. Let 1 s p1
. If
of negative residue classes among a, 2a, . . . , p1
2
2
p1
sa
sa mod p is a positive residue class we write sa = [ p ]p + us with 1 us 2 . If
sa mod p is a negative residue class we write sa = [ sa
]p+pus with 1 us p1
.
p
2
p1
A straightforward check shows that {u1 , u2 , . . . , u p1 } = {1, 2, . . . , 2 }. Addition
2
of these equalities yields
]
2 [
sa
p1
p1
sa = p
s=1
s=1
p1
+ p +
(us ).
s=1
Take both sides modulo 2,

p1
p1
s S(a, p) + +
s=1
us (mod 2)
s=1
p1
S(a, p) + +
s (mod 2).
s=1
The summations
on both sides cancel and we are left with S(a, p) (mod 2)
( )
a
2
hence p = (1) = (1)S(a,p) .
Theorem 5.2.6 (Quadratic reciprocity law) Let p, q be two odd prime numbers. Then
( )( )
p1 q1
p
q
= (1) 2 2 .
q
p
( )
( )
Alternatively pq = pq unless p q 1(mod 4), in which case we have
( )
( )
p
q
=
.
q
p
5.3. A GROUP THEORETIC PROOF
51
Proof. Let S(a, p) be as in Lemma 5.2.5. Then we assert

S(q, p) + S(p, q) =
p1q1
.
2
2
To see this, picture the rectangle [0, p/2] [0, q/2] and the lattice points (m, n)
N2 with 1 m p1
, 1 n q1
inside it. The diagonal connecting (0, 0) and
2
2
(p/2, q/2) does not pass through any of the lattice points. Notice that the number
of lattice points below the diagonal is precisely S(q, p) and above the diagonal
q1
lattice points, hence our assertion follows.
S(p, q). In total there are p1
2
2
We can now combine our assertion with Lemma 5.2.5 to obtain
( )( )
p1 q1
p
q
= (1)S(p,q)+S(q,p) = (1) 2 2 .
q
p
2
5.3
A group theoretic proof
It is known that Gauss gave six (more or less) dierent proofs of the quadratic
reciprocity law. Since then the number of proofs has increased dramatically to an
estimated 200. The proof we have given above is essentially due to Eisenstein. In
P.Bachmann, Die Lehre von der Kreistheilung, de Gruyter, Berlin, Leipzig, 1921
we nd several classical proofs and we can nd another 25 in O.Baumgart, Uber

das quadratische Reciprocitatsgesetz, Teubner, Berlin, 1885. In the article On
the quadratic reciprocity law in J.Australian Math. Soc. 51(1991), 423-425 by
G.Rousseau we nd a proof of the reciprocity law which is surprisingly simple if
one is acquainted with elementary group theory. It turns out to be an application
of the chinese remainder theorem and we like to present it here.
Another proof of Theorem 5.2.6. Let notations be as in the theorem. We
work in the group G = ((Z/pZ) (Z/qZ) )/U where U = {(1, 1), (1, 1)}.
Clearly {(i, j) | i = 1, 2, . . . , p 1; j = 1, 2, . . . , (q 1)/2} is a full set of representatives of G. Their product equals
((p 1)!(q1)/2 , ((q 1)/2)!p1 ).
Since ((q 1)/2)!2 (1)(q1)/2 (q 1)!(mod q) we get
((p 1)!(q1)/2 , (1)
q1 p1
2
2
(q 1)!(p1)/2 ).
Another full set of representatives of G is given by {(k(mod p), k(mod q)) | k =

1, 2, . . . , (pq 1)/2; (k, pq) = 1}. This is a consequence of (Z/pqZ) (Z/pZ)
(Z/qZ) (chinese remainder theorem). The product of these elements modulo p
equals
(p1
) (p1)/2 q1
q3
( 2 p + i)
i=1 i(p + i)(2p + i) ( 2 p + i)
i=1
1 q 2q p1
q
2
52
which equals
(p 1)!
(q1)/2
/q
(p1)/2
(p 1)!
(q1)/2
( )
q
(mod p).
p
Similarly we compute the product modulo q and we obtain

( ))
(
( )
p
q
(p1)/2
(q1)/2
, (q 1)!
.
(p 1)!
p
q
Comparison of the two expressions for yields
(
) (( q ) ( p )) ( ( q ) ( p ))
p1 q1
1, (1) 2 2
,
1,
p
q
p
q
2
and hence the reciprocity law.
5.4
Applications
Example 1. Is x2 84(mod 97) solvable? Notice that

( ) ( )( )( ) ( )( ) ( )( )
84
4
3
7
97
97
1
1
=
=
=
= 1 1 = 1.
97
97
97
97
3
7
3
7
Hence our congruence equation is not solvable.
Example 2. Is 3x2 + 4x + 5 0 mod 76 solvable? According to the Chinese
remainder theorem this congruence is equivalent to
3x2 + 4x + 5 0(mod 4)
3x2 + 4x + 5 0(mod 19).
The rst equation is equivalent to x2 1(mod 4), which is solvable. Multiply the
second by 13 on both sides to obtain x(2 +14x
+8
)
( ) 0(mod 19). After splitting o
3
squares, (x + 7)2 3(mod 19). Since 19
= 19
= 1, the second congruence
3
equation, and hence the original one, is not solvable.
Example 3. Let p be an odd prime. Then,
( ) {
3
1
if p 1(mod 3)
=
1 if p 1(mod 3).
p
This follows from
( ) ( )( )
(p) (p)
p1
p1
3
1
3
=
.
=
= (1) 2 (1) 2
p
p
p
3
3
Since 1 is a quadratic residue modulo 3 and 1 a quadratic nonresidue our
assertion follows.
5.4. APPLICATIONS
53
Example 4. Let En be the integer whose digits in base 10 consist of n ones,

e.g. E13 = 1111111111111. These numbers are known as repunits. Alternatively
En = (10n 1)/9. As an example
( 10 )
(we
)like
( 5 )to show
( 67here
) that
(2E
) 33 is divisible by
2
67. We easily verify that 67 = 67 67 = 5 = 5 = 1. Hence, by
Theorem 5.1.3(b), 1033 1 mod 67 and hence 67|E33 .
Extensive calculations show that among the numbers En with n < 50000 only
E2 , E19 , E23 , E317 , E1031
are prime and E49081 is probably prime (H.Dubner, 1999). Here are some factorisations,
E3
E5
E7
E11
E13
E17
=
=
=
=
=
=
111 = 3 37
41 271
239 4649
21649 513239
53 79 265371653
2071723 5363222357
Theorem 5.4.1 Let p be a prime such that p 1(mod 4) and 2p + 1 prime.

Then 2p + 1 divides 2p 1.
(
)
2
Proof. Note that 2p + 1 is a prime which is 7(mod 8). Hence, 2p+1
= 1.
(
)
2
Theorem 5.1.3(b) now implies 2p 2p+1
1(mod 2p + 1).
2
As a corollary we see that the Mersenne numbers
223 1, 283 1, 2131 1
are not prime. For p < 10000 there are 100 values for which Theorem 5.4.1
applies. (See also Theorem 5.4.3 below and the chapter on applications of residues
for Mersenne numbers). We also note that if p is prime, then any prime divisor q
of 2p 1 has the form q = 2pk + 1. So, when looking for prime divisors of 2p 1
it makes sense to start by trying 2p + 1.
Theorem 5.4.2 (P
epin, 1877) For any n N let Fn = 22 + 1. Then,
n
Fn is prime 3 2 (Fn 1) 1(mod Fn ).

1
Proof. Because Fn is an odd prime we have

( ) ( ) ( )
1
3
Fn
1
(F
1)
n
1(mod Fn ).
32
Fn
3
3
54
The second to last congruence follows from 22 + 1 (1)2 + 1 1 + 1

1(mod 3).
n
Notice that Fn 1 = 22 and 3Fn 1 1(mod Fn ). Hence ord(3) divides
n
22 and equals 2r for some 0 r 2n . Suppose r < 2n then we would have
3(Fn 1)/2 1(mod Fn ), contradicting our assumption. Hence r = 2n and ord(3) =
Fn 1. In general, if we have a Z such that ord(a) in (Z/mZ) is m 1 then
m must be prime. In particular, Fn is prime.
2
n
The numbers Fn are known as the Fermat numbers, see the chapter on applications of congruences for more on the primality of Fn .
Theorem 5.4.3 (Lucas, Lehmer) For any n N let Mn = 2n 1. Dene
S1 , S2 , . . . by the recursion
S1 = 4
Sk+1 = Sk2 2,
k 0.
If n 3 and odd then,

Mn is prime Mn divides Sn1 .
Inthe proof of Lucas criterion we shall work in rings of the form Rm := {a +
b 3| a, b Z/mZ} where m N. More strictly speaking, Rm := Z[X]/(m, X 2
3) or Z[ 3]/(m). We state one result in Rm separately.
Let p be a prime larger than 3. Let a, b Z. Then, (a + b 3)p

Lemma
( ) 5.4.4
a + p3 b 3(mod p).
Proof. Since p is a prime we have, using Fermats little Theorem 3.3.5,
(a + b 3)p ap + bp ( 3)p (mod p)

p1
a + b 3 2 3(mod p).
Our lemma now follows if we use Eulers Theorem 5.1.3(b).
Proof of Theorem 5.4.3. By induction on k it is not hard to show that
k1
k1
k 1.
Sk = (2 + 3)2 + (2 3)2
The condition Mn |Sn1 can be rewritten as follow,
n2
n2
Sn1 0(mod Mn ) (2 + 3)2 + (2 3)2
0(mod Mn )
n1
(2 + 3)2 + 1 0(mod Mn )
n1
(2 + 3)2
1 (mod Mn ).
5.5. JACOBI SYMBOLS, COMPUTING SQUARE ROOTS
55
2n2
The second equivalence
follows
by
multiplication
with
(2
+
3)
and using the
fact that (2 + 3)(2 3) = 1. The proof of our theorem now comes down to
proving that
n1
Mn is prime (2 + 3)2
1(mod Mn ).
The latter congruence can also be written as (2 + 3)(Mn +1)/2 1(mod Mn )

and we can now note the similarity with Pepins test.
. First notice that, n being odd and 3, we have Mn 7(mod 24). Hence
if Mn is prime then 3 is a quadratic nonresidue mod Mn and 2 is a quadratic
residue mod Mn . Since Mn is assumed prime we have according to Lemma 5.4.4,
(
)
M
3
n
3 1 3(mod Mn ).
(1 + 3) 1 +
Mn
Thus we nd,
n
(1 + 3)2 (1 + 3)Mn (1 + 3) (1 3)(1 + 3) 2(mod Mn ). (5.1)
On the other hand,
n
n1
n1
n1
n1
(1+ 3)2 (1+ 3)22
(4+2 3)2
22 (2+ 3)2 (mod Mn ). (5.2)
Since 2 is a quadratic residue modulo Mn , we have
n1
22
2 2(Mn 1)/2 2(mod Mn ).
(5.3)
Combination of (5.1), (5.2), (5.3) nally yields

n1
(2 + 3)2
1(mod Mn )
as asserted.
. Let p be a prime divisor of Mn such that p 1(mod 12). Since Mn
7(mod 12) such a p exists. In particular we
that 3
is a quadratic nonresidue
have
p
mod p. Hence Lemma
5.4.4
yields
(2
+
3)
3(mod p). After multi

p+1
plication by 2 + 3 we nd (2+ 3)
1(mod p). On the other hand,
by
2n1
1(mod p). This implies that 2 + 3
assumption we have that (2 + 3)
has order 2n in the unit group of Z[ 3]/(p). Hence 2n divides p+1. But p divides
Mn = 2n 1. Thus we conclude that p = Mn and Mn is prime.
2
5.5
Jacobi symbols, computing square roots
( 111 )
To determine the Legendre symbol 137
say, we must rst factor 111 before being
able to apply quadratic reciprocity. This( is all right)for small numbers like 111,
11111111111
? (197002597249 is prime)
but what to do if we want to compute 197002597249
56
or Legendre symbols with even larger numbers? In the chapter on applications

of congruences we pointed out that factorisation of large numbers is a major
computational problem. Luckily this does not mean that the computation of
Legendre symbols becomes dicult. The solution is to use the slightly more
general Jacobi symbol.
Definition 5.5.1 Let n N be odd and m Z such that (m,(n)) = 1. Let
n = p1 p2 pr be the prime factorisation of n. The Jacobi symbol m
is dened
n
by
(m) (m) (m) (m)
=
n
p1
p2
pr
( )
where the symbols m
are the Legendre symbols.
pi
( )
Remark 5.5.2 Note that if m
= 1 then x2 m(mod n) is not solvable
n
2
simply
( m ) because x m(mod pi ) is not solvable for some i. 2On the other hand,
anything about the solubility of x m(mod n). For
if n = (1 we
) cannot say
1
2
example, 21 = 1 but x 1(mod 21) is certainly not solvable.
However, we do have the following theorem.
Theorem 5.5.3 Let n, m be odd positive integers. Then,
i)
ii)
iii)
1
n
)
= (1)
n1
2
( )
n2 1
2
= (1) 8
n
(m) ( n )
n
= (1)
m1 n1
2
2
Proof. These statements can be proved by using the corresponding theorems for
the Legendre symbol and the observation that for any r-tuple of odd numbers
u1 , . . . , ur we have
ur 1
u 1 ur 1
u1 1
+ +
(mod 2).
2
2
2
(5.4)
To be more precise, the sum on the left of (5.4) is modulo 2 equal to the number k
of ui which are 1(mod 4). If k is even, the product u1 ur is 1(mod 4) and the
term on the right of (5.4) is also even. If k is odd, we have u1 ur 1(mod 4),
hence the term on the right of (5.4) is also odd.
5.5. JACOBI SYMBOLS, COMPUTING SQUARE ROOTS
57
Let n = p1 pr be the prime factorisation of n. Then i) follows from

( ) ( ) ( )
p1 1
pr 1
1
1
1
=
= (1) 2 ++ 2
n
p1
pr
and
p1 1
pr 1
p1 pr 1
n1
+ +
(mod 2).
2
2
2
2
Similarly, ii) follows from
( ) ( )
( )
p2
p2
2
2
2
r 1
1 1
=
= (1) 8 ++ 8
n
p1
pr
and
p21 1
p2 1
(p1 pr )2 1
n2 1
+ + r
(mod 2).
8
8
8
8
Let m = q1 qs be the prime factorisation of m. Then iii) follows from
(m) ( n ) ( q ) (p )
pj 1 qi 1
i
j
=
= (1) i,j 2 2
n
m
pj
qi
i,j
and
p j 1 q i 1 q i 1 pj 1
m1n1
(mod 2).
2
2
2
2
2
2
i,j
i
j
2
( 11111111111 )
The computation of 197002597249 can now be done using a euclidean-like algorithm and Theorem 5.5.3. Notice that
197002597249 = 17 11111111111 + 8113708362
8113708362 = 2 4056854181
11111111111 = 2 4056854181 + 2997402749
...
...
Hence,
(
) (
)
11111111111
197002597249
=
197002597249
11111111111
(
) (
)(
)
8113708362
2
4056854181
=
=
11111111111
11111111111
( 11111111111 )
11111111111
=
=
4056854181
We keep repeating these steps of inversion and extraction of factors 2 until we
nd the value of the Jacobi symbol to be 1. From this algorithm we see that
58
computation of Jacobi symbols, and hence Legendre symbols, is polynomial in

the length of the input.
Let)p be a prime and a Z. Suppose that by compution of the Legendre symbol
(
a
we know that x2 a(mod p) is solvable. How do we nd a solution? The
p
following algorithm is very fast, provided we have a quadratic nonresidue mod
p at our disposal. Assuming the truth of the so-called Generalised Riemann
Hypothesis there exists a quadratic nonresidue between 1 and 2(log p)2 and the
algorithm we describe is then polynomial in the length of the input. Although we
do not know for sure that there exist small quadratic nonresidues , the Riemann
hypothesis is still unproved, experience seems to conrm their existence. So in
practice the algorithm described here is quite fast.
Solution of x2 a(mod p), Tonellis algorithm. Let p 1 = 2s m, with m
odd. Since (Z/pZ) is a cyclic group the group G of elements whose order divides
2s is also cyclic. To nd a generator of this group we must determine a quadratic
nonresidue modulo p. This can be done by just trying 2, 3, 5, 7, . . .. One very soon
hits upon a quadratic nonresidue . Call it x0 . We assert that b xm
0 (mod p) is
p1
2s
m2s
a generator of G. First of all note that b x0
x0 1(mod p), hence
s
the order of b divides 2s . On the other hand, if the (order
) of b is less than 2 ,
then b is an even power of the generator of G, hence pb = 1. This contradicts
( ) ( )m
b
= xp0
= (1)m = 1. Hence b generates G.
p
Let x be a solution of x2 a(mod p). Raise both sides to the power (m + 1)/2,
x xm a(m+1)/2 (mod p). Note that xm G. Hence, up to an element of G,
the solution x equals a(m+1)/2 (mod p). So we simply try a(m+1)/2 bj as a solution
for j = 0, 1, . . . , 2s1 1. So, if 2s is small we nd a solution very soon by just
trying j = 0, 1, . . . , 2s1 1. If 2s is large we can use an additional trick, due to
D.Shanks, which is described below.
As an example we solve x2 11111111111(mod 197002597249). Write p =
197002597249. It turns out that 7 is the least quadratic nonresidue mod p. We
have p 1 = 27 1539082791. Let m = 1539082791. Then,
b 7m 59255134607(mod p)
r 11111111111(m+1)/2 68821647300(mod p).
All we have to do now is try the numbers bj r(mod p) with j = 0, 1, 2, . . . , 63 as
solution of x2 11111111111(mod p). It turns out that j = 37 does the job and
we nd
574553913082 11111111111(mod 197002597249).
If, in Tonellis algorithm the value of 2s is large, we can speed up the last part of
the algorithm as follows (Shanks),
Let G be a cyclic group of order 2s with generator b. Let g G. Then the output
e of the following algorithm is precisely the number such that g = be .
5.6. CLASS NUMBERS
59
e := 0
loop:
Choose t Z0 minimal such that g 2 = 1.
t
If (t > 0) g := gb2
st
, e := e + 2st , goto loop
If (t == 0) stop
By using this algorithm with g = x2 am1 we can determine the value of 2j in
Tonellis algorithm quickly.
5.6
Class numbers
Take an odd prime p > 3 and consider the sum of quadratic residues
a.
R=
1ap1,a residue mod p
Let N be the analogous sum of quadratic non-residues. Notice that
(p1)/2
k=1
1
k 2 p(p 1)(2p 1) 0(mod p).
6
Also,
R+N
p1
1
a p(p 1) 0(mod p).
2
a=1
Hence both R and N are divisible by p. Let us make a table of (N R)/p.

p
(N R)/p
5 7
0 1
11
1
13
0
17
0
19
1
23
3
29
0
31
3
37
0
41
0
43
1
47
5
53
0
59
3
This table suggests that (N R)/p is always 0 and 0 if and only if p 1(mod 4).
Suppose rst that p 1(mod 4). If a is a residue modulo p, the same holds for
p a. So the quadratic residues mod p come in pairs and there are (p 1)/4
such pairs. Moreover, the sum of each pair is p, hence R = p(p 1)/4. The
same argument shows that N = p(p 1)/4. This conrms our expectation that
N R = 0 if p 1(mod 4). Proving that (N R)/p > 0 if p 3(mod 4)
is far more dicult however. A well-known proof by Dirichlet uses arguments
from complex function theory. Here is a table of values when p 3(mod 4) and
p < 200.
p
(N R)/p
7 11
1 1
19
1
23
3
31
3
43
1
47
5
59
3
67
1
71
7
79
5
83
3
103
5
60

p
107
(N R)/p 3
127
5
131
5
139
3
151
7
163
1
167
11
179
5
191
9
199
9
When
[ p ] 3(mod 4) we call (N R)/p the class-number of the ring O =
Z 1+ 2 p . Notation: h(p). These class numbers form the tip of an iceberg,
which is the eld of quadratic forms, arithmetic in quadratic elds and Dirichlet L-series. For example, the interest of this class number lies in the fact that
h(p) = 1 if and only if we have unique factorisation in irreducibles in O. It was
already suspected by Gauss that the biggest p for which h(p) = 1 is p = 163.
This was only proved in the 1950s by Heegner, Stark and later, in the 1960s via
other methods, by A.Baker.
5.7
Exercises
Exercise 5.7.1 Determine the quadratic residues modulo 17 and 19.

Exercise 5.7.2 Fermat observed that if a, b N and (a, b) = 1 and p is an odd
prime divisor of a2 + b2 , then p 1(mod 4). Prove this.
Exercise 5.7.3 Let p be an odd prime and a Z not divisible by p. Prove by
induction on k,
x2 a(mod p) has a solution x2 a(mod pk )has a solution.
Exercise 5.7.4 Let p be a prime and suppose p 1(mod 8).
a) Prove that x4 1(mod p) has a solution.
b) Choose a solution x of x4
p) and compute the residue class (x +
( 1(mod
)
2
1 2
x ) (mod p). Conclude that p = 1 if p 1(mod 8).
Exercise 5.7.5 Let p be a prime and suppose p 1(mod 3).
1. Prove that x2 + x + 1 0(mod p) has a solution.
( )
2. Prove, using a), that 3
= 1 if p 1(mod 3).
p
3. Determine the discriminant of x2 + x + 1.
(
4. Prove,using considerations such as in a) and b) that
1(mod 3).
Exercise 5.7.6 Show that
( 105 )
131
3
p
= 1 if p
= 1 and solve x2 105(mod 131).
5.7. EXERCISES
61
Exercise 5.7.7 Verify which of the following equations are solvable,

x2
x2
x2
x2
x2
9x2 + 12x + 15
8x2
114(mod 127)
61(mod 93)
47(mod 101)
47(mod 143)
837(mod 2996)
0(mod 58)
2x + 3(mod 175).
Exercise 5.7.8 For which prime numbers is 5 a quadratic residue ? Same question for 3 and 3.
Exercise 5.7.9 Let a Z and p, q two odd (
primes
) (not) dividing a. Prove,
a
a) If a 1(mod 4) then: p q(mod |a|) p = aq .
( ) ( )
b) If a 1(mod 4) then: p q(mod 4|a|) ap = aq .
Exercise 5.7.10 Let p be a prime which is 1 modulo 4. Let a be a quadratic
residue modulo p. Prove that a solution of x2 a(mod p) is given by x = a(p+1)/4 .
Exercise 5.7.11 Let p be a prime which is 5 modulo 8. Let r be an element
of (Z/pZ) of order 4 and let a be a quadratic residue modulo p. Prove that a
solution of x2 a(mod p) is given by either x = a(p+3)/8 or x = ra(p+3)/8 .
Exercise 5.7.12 Prove that there exist innitely many primes p such that p
1(mod 4). Prove that there exist innitely many primes p such that p 1(mod 4).
Exercise 5.7.13 Let p be an odd prime. Let a be the smallest positive integer
such that a is a quadratic nonresidue modulo p. Show that a < 1 + p.

Exercise 5.7.14 a) Let p = 4k + 1 be prime. Prove,
p2 1
[ p] + [ 2p] + + [ kp] =
.
12
b) Let p = 4k + 3 be prime. Prove,
(p 1)(p 2)
.
[ p] + [ 2p] + + [ kp]
12
(Hint: count the number of lattice points below the parabola y = px with x < p/4
and use in b) that N K.)
Chapter 6
Dirichlet characters and Gauss
sums
6.1
Characters
In this chapter we study an important tool in number theory, namely characters

on (Z/mZ) .
Definition 6.1.1 Let G be a nite abelian group. A homomorphism : G C
is called a character of G.
Since any element g of a nite group has nite order, (g) must have nite order
in C , hence be a root of unity. In particular, (e) = 1 for all . Moreover if G
is cyclic and g is a generator of G, then is determined by its value (g). Any
other a G is of the form a = g k and so we must have (a) = (g k ) = ((g))k .
Definition 6.1.2 The trivial character or principal character on a nite abelian
group G is the character given by (g) = 1 g G. Notation: 0 .
Suppose we have two characters 1 and 2 on G. Then one easily veries that the
new function 1 2 given by (1 2 )(g) = 1 (g)2 (g) g G is again a character
on G. In fact, the characters on G form a group with the trivial character as
Correspondingly we can speak
identity element. We denote this group by G.
about the order of a character as being its order in G.

Lemma 6.1.3 Let G be a nite abelian group and a non-trivial character on
G. Then,
(g) = 0.
gG
62
6.1. CHARACTERS
63
Proof. Since is non-trivial, there exists h G such that (h) = 1. We now

use the fact that if g runs through G then so does gh,
(g) =
(hg) = (h)
(g).
gG
gG
gG
Because (h) = 1 our assertion follows.
To get a good grasp on the characters of an abelian group we consider the vectorspace V = C|G| consisting of |G|-tuples of complex numbers (cg1 , cg2 , . . .) indexed by the elements g1 , g2 , . . . of G. On V we dene a complex inner product
by
(x, y ) =
xg y g
gG
for any x = (xg1 , xg2 , . . .), y = (yg1 , yg2 , . . .) in V . The bar denotes complex
conjugation. We let any h G act on V by
hx = (xhg1 , xhg2 , . . .).
So we see that any h simply permutes the coordinates of x in a way prescribed
by the groupstructure. In particular, the elements of G act as unitary (=length
preserving) linear maps on V . Let be a character on G. We dene the corresponding charactervector v by
v = ((g1 ), (g2 ), . . .)
in other words, v is just the sequence of values of . Notice that for any h G,
hv = ((hg1 ), (hg2 ), . . .) = (h)((g1 ), (g2 ), . . .) = (h)v .
Hence v is a common eigenvector for all h G with eigenvalues (h). Conversely, any common eigenvector v of all h G denes a character on G by
taking the eigenvalue of each h as charactervalue. Let hv = (h)v and suppose we had taken g1 = e. Then h(ve , . . .) = (vh , . . .) and, on the other hand,
h(ve , . . .) = (h)(ve , . . .) = ((h)ve , . . .). Hence vh = (h)ve for all h G. So
v = ve ((g1 ), (g2 ), . . .). In particular, v is uniquely dened up to a scalar factor.
Finally, let 1 , 2 be two distinct characters. In particular, 1 2 = 1 1
2 is not
the trivial character. Then, using Lemma 6.1.3,
(v1 , v2 ) =
1 (g)2 (g) =
(1 2 )(g) = 0,
gG
gG
hence v1 and v2 are orthogonal.

= |G|.
Lemma 6.1.4 |G|
64
CHAPTER 6. DIRICHLET CHARACTERS AND GAUSS SUMS
Proof. Since, by the above remarks, character vectors belonging to distinct characters are orthogonal, there cannot be more than |G| characters. Suppose that we
have k distinct characters and that k < |G|. Then the orthogonal complement U
in V of the character vectors has dimension |G| k > 0. Moreover, every element
h G is unitary and hence maps U into itself. From linear algebra we know
that any set of commuting unitary operators has a common eigenvector, which
in its turn denes a character of G, distinct from the ones we already had. So
we now have k + 1 distinct characters. We continu this procedure untill we have
|G| distinct characters.
2
Lemma 6.1.5 Let G be a nite abelian group and g G not the identity element.
Then,
(g) = 0.
= {1 , 2 , . . .}. Consider the |G| by |G|

Proof. Let G = {g1 , g2 , . . .} and G
matrix M = (i (gj ))i,j=1,...,|G| . We have seen in the remarks above that the rows
(letting
j run) of this
matrix are orthogonal. Moreover, all rows have lengths
|G|. Hence M is |G| times a unitary matrix. In such a matrix the columns
are orthogonal as well, in other words for any distinct g, h G we have
(g)(h) = 0.
Our Lemma follows by taking h = e.
Definition 6.1.6 A Dirichlet character is a character on (Z/mZ) .

Remark 6.1.7 Let be a Dirichlet character on (Z/mZ) . Often one extends
to a function on Z by taking the value (n) for all n with (n, m) = 1 and
(n) = 0 whenever (n, m) > 1. By abuse of language we then still speak of a
Dirichlet character.
Examples
( )
i. The Legendre symbol p. on (Z/pZ) for odd primes p and the Jacobi
symbol on (Z/mZ) for odd m. Both examples are real characters of order
two.
ii. The character of order four on (Z/13Z) given by
a: 1 2 3
4 5 6 7 8 9 10 11 12 (mod 13)
(a) : 1 i 1 1 i i i i 1 1 i 1
One easily veries that (ab) = (a)(b) for all a, b (Z/13Z) . We have
constructed this character by assigning (2) = i. Since 2 is a primitive root
modulo 13, the powers of 2 run through (Z/13Z) and we can determine
the character values correspondingly.
6.2. GAUSS SUMS, JACOBI SUMS
65
iii. The non-trivial characters 1 , 2 , 3 on (Z/8Z) given by

a:
1 (a) :
2 (a) :
3 (a) :
1
3
5
7
1 1
1 1
1 1 1
1
1
1 1 1
Definition 6.1.8 A Dirichlet character is called even if (1) = 1 and odd if

(1) = 1.
6.2
Gauss sums, Jacobi sums
Although it is possible to consider more general Gauss sums, we shall restrict

ourselves here to (Z/pZ) where p is an odd prime. For the p-th roots of unity
we mention the following Lemma. In the sequel we let p denote e2i/p .
Lemma 6.2.1 Let n Z and p an odd prime. Then,
p1
{
pnx
x=1
1
if p/
|n
p 1 if p|n
Proof. Clearly, when p|n all terms in the summation are 1, so the second case
follows immediately. In the rst case we use the summation formula for geometric
sequences to obtain
(ppn pn )/(pn 1) = 1.
Definition 6.2.2 Let p be an odd prime and a Dirichlet character on (Z/pZ) .

The Gauss sum S is dened by
S =
p1
(x)px .
x=1
Theorem 6.2.3 Let be a non-trivial Dirichlet character on (Z/pZ) . Then,

a) S S1 = (1)p,
b) |S |2 = p,
c) if is the Legendre symbol then S2 = (1)
p1
2
p.
66
Proof. Part a) We have

S S1 =
p1
(x)(y)1 px+y .
x,y=1
Replace x by zy and notice that if z runs through (Z/pZ) then so does x = zy

for any xed y (Z/pZ) . Hence
S S1 =
p1
(z)(y)(y)1 py(z+1) .
z,y=1
The factors (y) cancel. According to Lemma 6.2.1 summation over y now yields
1 if z 1(mod p) and p 1 if z 1(mod p). Hence,
S S1 = (1)p
p1
(x).
x=1
The summation on the right vanishes according to Lemma 6.1.3 and we have
proved assertion a).
Part b) First of all, notice that
S1 =
p1
(x)1 px
x=1
p1
(x)
(x)p
x=1
= (1)
p1
(x)
(x)p
x=1
= (1)S .
Hence |S |2 = S S = (1)S S1 and assertion b) follows by using assertion
a).
c) If is the Legendre symbol we have = 1 and hence, via assertion a),
2
S2 = (1)p = (1)(p1)/2 p.
Definition 6.2.4 Let 1 , 2 be Dirichlet characters modulo p, where p is an odd
prime. The Jacobi sum J(1 , 2 ) is dened by
J(1 , 2 ) =
p1
x=2
1 (x)2 (1 x).
6.3. APPLICATIONS
67
Theorem 6.2.5 Let notations be as in Denition 6.2.4. Suppose that 1 and 2

are not each others inverse. Then,
S1 S2 = J(1 , 2 )S1 2 .
Proof. Write
S1 S2 =
p1
1 (x)2 (y)px+y .
x,y=1
We introduce the new summation variables u, v via x = uv, y = u(1 v).

Conversely, u = x + y, v = x/(x + y). One easily veries that this gives us a
bijection between the sets {(x, y) (Z/pZ) 2 | x + y 0(mod p)} and {(u, v)
(Z/pZ) 2 | v 1(mod p)}. Hence,
S1 S2 =
1 (x)2 (y)p(x+y) +
p1 p1
xy
u=1 v=2
p1 p1
p1
1 (uv)2 (u(1 v))pu
1 (1)1 (x)2 (x) +
x=1
1 (v)2 (1 v)1 (u)2 (u)pu .
u=1 v=2
The rst summation vanishes because of Lemma 6.1.3 and the fact that 1 2 =
0 . In the second summation the sum over u yields S1 2 and summing over v
yields J(1 , 2 ). Hence S1 S2 = J(1 , 2 )S1 2 .
2
Corollary 6.2.6 Let notations be as in Denition 6.2.4. Suppose that the char
acters are not each others inverse. Then |J(1 , 2 )| = p.
Proof. This follows from Theorem 6.2.5 and Theorem 6.2.3(b).
6.3
Applications
A nice application of Gauss sums is a short proof of the quadratic reciprocity law.
Let p, q be two odd primes. Write p = S where is the Legendre symbol on
(Z/pZ) . The following calculations will be performed in Z[p ] considered modulo
q. Using Theorem 6.2.3(c) and Eulers Theorem 5.1.3(b) we nd that
( )
p1
q1
p1 q1
p
q1
q
p p p ((1) 2 p) 2 p (1) 2 2
p (mod q).
q
On the other hand,
pq
( p1 ( )
x
x=1
)q
px
68
p1 ( )
pqx
( )
)
p1 (
q
qx
pqx
p x=1 p
p
x=1
(
)
q
p (mod q).
p
Comparison of the above two equalities yields

( )
( )
p1 q1
q
p
p
p (mod q).
(1) 2 2
q
p
Multiply on both sides by p and cancel the resulting p2 = (1)
obtain
( ) ( )
p1 q1
q
p
(1) 2 2
(mod q)
q
p
p1
2
p. We then
and since q 3, strict equality follows.
( )
Using the above idea we can also compute p2 . The following calculation will
be performed in Z[i] modulo p. We consider (1 + i)p (mod p) and compute it in
two ways. First of all
{
p1
1 + i(mod p)
if p 1(mod 4)
p
p
(1 + i) 1 + i 1 + (1) 2 i
i(1 + i)(mod p) if p 3(mod 4)
On the other hand, since (1 + i)2 = 2i,
(1 + i) (2i)
p
p1
2
( )
2 p1
(i + 1)
i 2 (1 + i)(mod p).
p
Comparison of these two results shows that

( ) { p1
p1
2
i 2 (1) 4
if p 1(mod 4)
p1
p+1
p
i i 2 (1) 4 if p 3(mod 4)
( )
One now easily checks that
3(mod 8).
2
p
= 1 if p 1(mod 8) and
( )
2
p
= 1 if p
Theorem 6.3.1 Let p be a prime with p 1(mod 4). Then p can be written as
the sum of two squares. More precisely, p = a2 + b2 with
)
2 (
x(x2 1)
p1
a=
x=1
)
2 (
x(x2 )
p1
where is any quadratic nonresidue .

b=
x=1
6.3. APPLICATIONS
69
Remark 6.3.2 The sum

cobsthal sum.
p1 ( x(x2 1) )
x=0
, which equals 2a, is known as the Ja-
Proof. Since p 1(mod 4) we have a character of order 4 on (Z/pZ) which we

call 4 . This can be constructed by taking (g) = i for a primitive root g. Any
other element of a (Z/pZ) has the form a g k (mod p) and so( (a)
) is xed( by
)
.
k
k
k
(a) = (g ) = (g) = i . Consider the Jacobi sum J = J(4 , p ), where p.
is the Legendre symbol. Notice that J Z[i], i.e. J = a + bi, a, b Z. Moreover,
by Corollary 6.2.6, p = |J|2 = a2 + b2 . Hence our rst assertion is proved.
Notice that in the summation
J=
p1
(
4 (x)
x=1
1x
p
a term is real if x is a quadratic residue and purely imaginary if x is a quadratic

nonresidue . Hence
(
p1
a = J =
4 (y )
y=1
1 y2
p
)
2 ( )(
y
1 y2
p1
y=1
)
2 (
y(y 2 1)
p1
y=1
Taking to be any quadratic nonresidue we get

(
p1
ib = iJ =
4 (y )
y=1
1 y 2
p
)
2 ( )(
y
1 y 2
p1
= 4 ()
y=1
)
2 (
y(y 2 )
p1
= 4 ()
y=1
Since 4 () = i our asserted value of follows up to sign, which is sucient. 2

Another amusing application is the following theorem discovered by Gauss,
Theorem 6.3.3 Let p be an odd prime for which p 1(mod 3). Then,
x Z : x3 2(mod p) A, B Z : p = A2 + 27B 2 .
When p 1(mod 3), the congruence x3 2(mod p) always has a solution.
70
Proof. When p 2(mod 3) we simply take x = 2(2p)/3 as solution (verify!). So

we suppose that p 1(mod 3). Then there exists a character mod p of order
3. First of all we notice that
x Z : x3 2(mod p) (2) = 1.
Notice that the third roots of unity are all distinct modulo 2. Hence it follows
from Lemma 6.3.4 that
x Z : x3 2(mod p) J(, ) 1(mod 2).
We know that J(, ) is an element of Z[], the ring of Eisenstein integers, where
is a cube root of unity. Thus we have J(, ) = a + b for some a, b Z. From
Lemma 6.3.4 it follows that a 1(mod 3) and b 0(mod 3). Similarly, if
(2) = 1 we nd
that b 0(mod 2). When b iseven we can rewrite a + b as
a b/2 + (b/2) 3 hence it is an element of Z[ 3]. So we now nd
x Z : x3 2(mod p) A, B Z : J(, ) = A + 3B 3
where we have put A = a b/2, B = b/6.
Suppose that x3 2(mod p) has a solution. ThenJ(, ) = A + 3B 3 and

since |J(, )|2 = p, we nd that p = A2 + 27B 2 .
2
Suppose conversely that p can be written as p = A
+ 27B 2 . Then we have in
Z[] the factorisation p = where = A 3B 3. Choosing the proper
sign, we know, by unique factorisation, that equals J(, ) up to a unit. Since
1(mod 3) and J(,
) 1(mod 3) this unit must be 1 and hence J(, )
is of the form A + 3B 3. This implies that x3 2(mod p) is solvable.
Lemma 6.3.4 Let p be an odd prime satisfying p 1(mod 3). Let be a Dirichlet character mod p of order 3. Then,
J(, ) 1(mod 3)
J(, ) (2)(mod 2).
Proof. According to Theorem 6.2.5 we have

S3 = S S2 J(, ) = pJ(, ).
Notice that modulo 3 we have
S3
p1
(x)3 p3x (mod 3)
x=1
p1
p3x 1(mod 3)
x=1
where the last equality follows from Lemma 6.2.1.

6.4. EXERCISES
71
Notice that modulo 2 we have

S4
p1
(x)4 p4x (mod 2)
x=1
1
(4)
p1
(4x)p4x (mod 2)
x=1
(2)S (mod 2)
Our congruence mod 2 follows after multiplication by S , Theorem 6.2.3(a) and
p 1(mod 2).
2
6.4
Exercises
Exercise 6.4.1 Determine all Dirichlet characters modulo 7 and modulo 12.
Exercise 6.4.2 Choose a character 4 of order 4 and a character 2 of order 2
on (Z/13Z) . Give a table in which x, 2 (x), 4 (x) and 2 (x)4 (1 x) are listed
for x = 1, . . . , 12. Compute the Jacobi sum J(2 , 4 ). Note that it is a number
in Z[i] and that its norm is 13.
Exercise 6.4.3 Proof that a cyclic group of order N has precisely N characters,
which again form a cyclic group.
Exercise 6.4.4 Let p be an odd prime. Prove that there exists a Dirichlet character of order 4 if and only if p 1(mod 4).
Exercise 6.4.5 Let p be a prime and a Z, not divisible by p. Suppose p
2(mod 3). Prove that x3 a(mod p) has precisely one solution modulo p.
Exercise 6.4.6 Determine all solutions of x3 2(mod p) for p = 19, 31. Check
also if p can be written in the form p = a2 + 27b2 .
Chapter 7
Sums of squares, Warings
problem
7.1
Sums of two squares
In previous chapters we have already seen that if an odd prime divides the sum of
two relatively prime squares, it is 1(mod 4). Moreover, Fermat showed that any
prime 1(mod 4) can be written as a sum of two squares. Furthermore it was
observed since antiquity that any positive integer can be written as the sum of
four squares. This was proved by Lagrange for the rst time in 1770. Problems
such as these form the subject of this section. It should be noted explicitly here
that by a square we mean the square of a number in Z. So, 02 is also considered
to be a square.
The quickest and most elegant way to deal with the above mentioned problems
is to work in the rings of Gaussian integers and quaternionic integers where
we rely heavily on the fact that they are euclidean domains. There exist also
presentations of this subject which work entirely in Z. However, they are in fact
a disguised form of application of the euclidean algorithm.
Theorem 7.1.1 (Fermat) Let p be an odd prime. Then p is a sum of two
squares if and only if p 1(mod 4).
Proof. Suppose p = x2 + y 2 for some x, y Z. Since squares are either 0 or 1
mod 4, p, being odd, can only be 1(mod 4).
Now suppose that p 1(mod 4). Then the congruence equation x2 1(mod p)
has a solution, x0 say. Let us now work in Z[i] and use unique factorisation. We
have p|(x20 + 1) hence p|(x0 + i)(x0 i). If p were prime in Z[i] we would have
p|(x i) and, via complex conjugation, p|(x + i). Hence p|2i, which is impossible.
Hence p = in Z[i] with N , N > 1. Take norms on both sides, p2 = N N .
Since N , N > 1 this implies p = N , hence p can be written as a sum of two
squares.
2
72
7.1. SUMS OF TWO SQUARES
73
Theorem 7.1.2 Let r2 (n) be the number of solutions x, y Z to n = x2 + y 2 .

i. The function r2 (n)/4 is multiplicative.
ii. Let p be a prime. Then
k + 1
r2 (pk ) 0
=
4
1
1
if
if
if
if
p 1(mod 4)
p 3(mod 4), k odd
p 3(mod 4), k even
p=2
Proof. Part i). We note that there is a 1-1-correspondence between solutions of

n = x2 + y 2 and factorisations n = in Z[i], simply by taking = x + yi. Also
note that r2 (n)/4 is precisely the number of factorisations n = where we count
, i i, , i i as being the same., i.e. we count factorisations
modulo units.
Let m, n N and (m, n) = 1. Let mn = , Z[i]. Let = gcd(, m),
= gcd(, n). Note that , are unique up to units. Then = (up to
units) and mn = = . Hence every representation of mn as sum of
two squares corresponds to a pair of representations of m = and n = .
Conversely, every such pair gives rise to representation mn = . Hence
r2 (mn)/4 = (r2 (m)/4)(r2 (n)/4).
Part ii). Let p = 2. Since 2 has only one prime divisor, 1 + i, in Z[i] we have, up
to units, the unique factorisation
2k = (1 + i)k (1 i)k .
Hence r2 (2k )/4 = 1.
Let p 3(mod 4). Then p itself is prime in Z[i]. For k odd there is no factorisation of the form pk = , when k is even we have only pk = pk/2 pk/2 .
Let p 1(mod 4). Then p = , where , are distinct primes in Z[i]. For
the factorisation pk = we have, up to units, k + 1 choices, namely =
k , k1 , . . . , k . Hence r2 (pk )/4 = k + 1, as asserted.
2
In order to solve n = x2 + y 2 for given n there exist several possibilities. When n
is small it is best to solve it by trial and error. When n is larger, there are two
possibilities. For medium sized numbers with several prime factors we suggest the
following method. Let n = pk11 pkr r be the prime factorisation of n. Write each
k
factor pj j as a sum of two squares a2j + b2j . Let j be one the complex numbers
k
aj bj i. Let rj=1 i = = a + bi. Notice that n = rj=1 pj j = rj=1 N (j ) =
N () = a2 + b2 . Since we have dierent choices for each j we will get dierent
solutions to n = x2 + y 2 . We have seen that we get essentially all solutions in this
way. As an example we write 65 as sum of two squares. Notice that 65 = 5 13
74
CHAPTER 7. SUMS OF SQUARES, WARINGS PROBLEM
and 5 = 12 + 22 and 13 = 22 + 32 . The multiplications (1 + 2i)(2 + 3i) = 4 + 7i

and (1 2i)(2 + 3i) = 8 i produce the solutions 65 = 42 + 72 and 65 = 82 + 12 .
According to Theorem 7.1.2 there are 16 solutions and they can be obtained from
the two solutions we found by interchanging x, y and the substitutions x x,
y y.
When n is a large prime we use a more sophisticated method. Via the techniques
described in the chapter on quadratic reciprocity we rst solve l2 1(mod n)
and then determine a + bi = gcd(n, l + i) via the euclidean algorithm in Z[i]. We
assert that n = a2 + b2 .
7.2
Sums of more than two squares
Theorem 7.2.1 (Lagrange 1770) Every positive integer can be written as a

sum of four squares.
Proof. First of all notice the equivalenc of the statements
i. n is sum of four squares
ii. n is the norm of a quaternionic integer.
Going from i. to ii. is trivial. Suppose conversely that n = N for some
quaternionic integer . According to Theorem 13.4.2 there exists a unit such
that = p + qi + rj + sk for some p, q, r, s Z. Hence n = N = N =
p2 + q 2 + r2 + s2 .
Secondly, if n, m can be written as a sum of four squares then so can nm. This
follows from N N = N . Thus it suces to prove that any prime p can be
written as a sum of four squares. We assume that p 5 and proceed as follows.
First we show that there exist x, y Z such that x2 +y 2 +1 0(mod p). Consider
the sets {i2 Z/pZ| 0 i p1
} and {1 i2 Z/pZ| 0 i p1
}. Each set
2
2
p+1
consists of 2 distinct elements in Z/pZ. Hence these sets overlap, and there
exist x, y such that x2 1 y 2 (mod p), as desired.
By shifting mod p we can see to it that we have x, y Z such that x2 + y 2 + 1
0(mod p) and |x|, |y| < p/2. Application of Theorem 13.2.3 to the numbers p and
1 + xi + yj yields a common right-divisor q (the right-gcd) and , q
such that
= p + (1 + xi + yj).
(7.1)
In the remainder of the proof b|a means: c q such that a = cb. First of all,
|p N |p2 N = 1, p or p2 . Secondly, |(1 + xi + yj) N |(1 + x2 + y 2 )
N < 2 (p/2)2 + 1 < p2 . Hence N = 1 or p. Suppose N = 1. Then, according
to Theorem 13.4.4 is a unit. Multiply Eq.(7.1) on both sides from the right by
(1 xi yj) to obtain (1 xi yj) = (1 xi yj)p + (1 + x2 + y 2 ). Since
7.2. SUMS OF MORE THAN TWO SQUARES
75
p|(1 + x2 + y 2 ) we infer p|(1 xi yj) and hence p|(1 xi yj), which is clearly
impossible. Thus we conclude that N = p and p is a sum of four squares.
2
Not everyone may be charmed by quaternions and for this reason we give Lagranges original proof of Theorem 7.2.1 as well.
Proof. Just as in the previous proof we note that it suces to prove that every
prime p can be written as the sum of four squares. This is a consequence of
Eulers identity,
(a2 + b2 + c2 + d2 )(a2 + b2 + c2 + d2 ) =
= (aa + bb + cc + dd )2
+(ab ba + cd dc )2
+(ac bd ca + db )2
+(ad + bc cb da )2 .
Just as above, we can nd x, y such that x2 +y 2 +1 0(mod p) and |x|, |y| p/2.
We then have x2 + y 2 + 1 = m0 p and m0 ((p/2)2 + (p/2)2 + 1)/p < p.
Let m be the smallest positive integer such that mp can be written as the sum
of four squares. We will show that m > 1 leads to a contradiction. Suppose
mp = a2 + b2 + c2 + d2
and m > 1.
(7.2)
Choose A, B, C, D in the interval (m/2, m/2] such that

a A(mod m), b B(mod m), c C(mod m) d D(mod m).
Then,
A2 + B 2 + C 2 + D2 a2 + b2 + c2 + d2 mp 0(mod m).
Hence,
mr = A2 + B 2 + C 2 + D2 ,
r 0,
(7.3)
where r = (A2 + B 2 + C 2 + D2 )/m (4 (m/2)2 )/m = m. First we show that r =

0, m. If r = 0, then A = B = C = D = 0 and hence m2 |a2 +b2 +c2 +d2 = mp from
which follows m|p, contradicting 1 < m < p. If r = m then A = B = C = D =
m/2 and hence a b c d m/2(mod m). Notice that if x m2 (mod m)
then x2 ( m2 )2 (mod m2 ). And so, mp = a2 +b2 +c2 +d2 4(m/2)2 0(mod m2 ),
which implies that m2 |mp and thus m|p, again a contradiction. We conclude,
0 < r < m.
(7.4)
Notice that in the proof of (7.4) we have used the assumption m > 1. Now
multiply (7.2) and (7.3) and use Eulers identity to obtain
mp mr = (a2 + b2 + c2 + d2 )(A2 + B 2 + C 2 + D2 ) = 2 + 2 + 2 + 2
76
where
= aA + bB + cC + dD a2 + b2 + c2 + d2 0(mod m)
= aB bA + cD dC ab ab + cd cd 0(mod m)
= aC Ac + dB bD ac ac + bd bd 0(mod m)
= aD dA + bC cB ad ad + bc bc 0(mod m)
So m divides , , , and we nd rp = (/m)2 + (/m)2 + (/m)2 + (/m)2 .
We conclude that rp is a sum of four squares and 0 < r < m. This contradicts
the minimality of m. Hence we have m = 1, as desired.
2
For completeness we like to mention the following theorem,
Theorem 7.2.2 (Jacobi) Let n N. Then the number of solutions to
n = x21 + x22 + x23 + x24 ,
equals
x 1 , x2 , x3 , x4 Z
d
8
where the denotes summation over all d with d|n, 4/

|d.
As to the question whether positive integers can be written as a sum of three
squares, we can easily give innitely many counterexamples.
Theorem 7.2.3 A positive integer of the form 4l (8k + 7) cannot be written as a
sum of three squares.
Proof. We proceed by induction on a. First of all, a number which is 7(mod 8)
cannot be the sum of three squares, simply because the sum of three squares can
never be 7(mod 8). This can easily be veried using the fact that a square can
only be 0, 1 or 4(mod 8).
Suppose l 0 and suppose we proved our theorem for numbers of the form
4l (8k +7). Let n = 4l+1 (8k + 7) and suppose n = a2 +b2 +c2 . Since n 0(mod 4)
we have necessarily that a, b, c are all even. Hence n/4 = (a/2)2 + (b/2)2 + (c/2)2 ,
contradicting our induction hypothesis.
2
The converse statement is much harder to prove.
Theorem 7.2.4 (Gauss) Any positive integer not of the form 4l (8k + 7) can be
written as the sum of three squares.
Corollary 7.2.5 Any positive integer can be written as the sum of three triangular numbers.
( )
A triangular number is a number of the form n2 .
7.3. THE 15-THEOREM
77
Proof. Let n N. Write 8n+3 as the sum of three squares, 8n+3 = a2 +b2 +c2 .
Notice that a, b, c all have to be odd. Write a = 2p 1, b = 2q 1, c = 2r 1.
Then 8n + 3 = 4(p2 p) + 4(q 2 q) + 4(r2 r) + 3 and this amounts to
( ) ( ) ( )
p
q
r
n=
+
+
.
2
2
2
(Quoting Gauss: EU REKA : num = + + !!)
Gauss also proved results on the number of representations of a number as sum

of three squares. As a curiosity we mention
Theorem 7.2.6 Let p be a prime with p 3(mod 8). Then the number of
solutions to p = x2 + y 2 + z 2 , x, y, z Z equals
24
a
p a=1
p1
( )
a
.
p
You might recognize the class number h(p) (see subsection 5.6) in this theorem.
So the number of ways in which a prime p 3(mod 8) can be written as sum of
three squares equals 24h(p).
Notice that squares can be considered as quadrangular numbers and Theorem 7.2.1
can be described cryptically as n = 2 + 2 + 2 + 2. Of course we can generalise
this and speak of pentagonal,
( ) hexagonal,... numbers. The general form of the
k-gonal numbers is (k 2) n2 +n. We mention the following theorem, conjectured
by Fermat and proven by Cauchy.
Theorem 7.2.7 (Cauchy) Any positive integer can be written as the sum of k
k-gonal numbers.
7.3
The 15-theorem
Recently Conway and Schneeberger found a remarkable theorem on the representation of integer by quadratic form. An integral quadratic form F (x1 , . . . , xr )
in r variables is a homogeneous polynomial of degree 2 with integer coecients.
It is called even if the coecients of the terms xi xj with i = j are all even.
The form F is called positive denite if F (x1 , . . . , xr ) > 0 for all choices of
(x1 , . . . , xr ) = (0, . . . , 0).
We say that a quadratic form F represents an integer if there exist integers
x1 , . . . , xr such that n = F (x1 , . . . , xr ). For example, the form x21 + x22 + x23 + x24
represents all positive integers. In 1993 Conway and Schneeberger announced the
following remarkable theorem.
78
Theorem 7.3.1 (Conway, Schneeberger, 1993) An integral, even positive definite quadratic form represents all positive integers if and only if it represents
1, 2, 3, 5, 6, 7, 10, 14, 15.
So, if we want to show that every positive integer can be written in the form
2x2 + 2xy + y 2 + z 2 + u2 with x, y, z, u integers, all we have to do is check whether
each of the numbers 1, 2, 3, 5, 6, 7, 10, 14, 15 can be written in this way. A simple
check shows that this is indeed the case.
The proof of this theorem was quite complicated and never published. However,
in 2000 Manjul Bhargava found a simpler proof and a generalisation to all positive
quadratic forms known as the 290-theorem.
Theorem 7.3.2 (Bhargava, Hanke, 2005) An integral, positive denite quadratic
form represents all positive integers if and only if it represents
1, 2, 3, 5, 6, 7, 10, 13, 14, 15, 17, 19, 21, 22,
23, 26, 29, 30, 31, 34, 35, 37, 42, 58, 93, 110,
145, 203, 290.
7.4
Warings problem
Around 1770 Waring put forward the following question . Let k N and k 2.
Is there a number g(k) such that any positive integer is the sum of at most g(k)
non-negative k-th powers? About 140 years later, in 1909, Hilbert proved that
the answer is yes for any k 2. His method is quite complicated and we shall
not reproduce it here. Somewhat later Hardy and Littlewood developed a more
conceptual approach to the problem and showed, among other things, that there
exists g(k) 3 2k . From now on we shall denote by g(k) the smallest number
for which Warings problem is solvable. The method of Hardy and Littlewood,
and further modications and improvements of it, are known as circle methods,
which have now become a standard tool in number theory (see R.C.Vaughan,
The circle method). It is now known that g(2) = 4 (Lagrange,1770), g(3) = 9
(Wieferich,Kempner,1912), g(4) = 19 (Balusabramanian, Deshouillers, Dress,
1986), g(5) = 37 (Chen-Jingrun, 1964). For k 6 we have, if (3/2)k [(3/2)k ] >
1 (3/4)k then g(k) = [(3/2)k ] + 2k 2. It is very likely, but not quite proved
yet, that the inequality always holds.
Although g(3) = 9 it turns out that only 23 and 239 require 9 cubes, all other
numbers can be written as the sum of at most 8 positive cubes. In connection
with this we can dene G(k), the smallest integer such that any suciently large
integer can be written as the sum of at most G(k) k-th powers. Since innitely
many numbers are not the sum of three or less squares we have G(2) = 4. It is
also known that 4 G(3) 7. It is generally suspected that G(3) = 4 but no
7.4. WARINGS PROBLEM
79
one has any idea how to prove this. By a rened version of the circle method
I.M.Vinogradov was able to show that G(k) 6k(log k +4 +log 216) and Wooley,
in 1992 improved this to G(k) k log k(1 + o(1)). Special results are G(4) = 16,
G(5) 17 (conjectured: 6), G(6) 24 (conjectured: 9), G(7) 33 (conjectured:
8), G(8) 42 (conjectured: 32), G(9) 50 (conjectured: 13), G(10) 59
(conjectured: 12) (see Vaughan and Wooley, Number Theory for the Milennium
III p301-340 A.K.Peters, 2000).
A variation on Warings problem is the following easier problem. Let k N, k
2. Does there exist a number s such that any n N can be written in the form
n = xk1 xk2 xks ?
If such an s exists we denote its minimal value by v(k). Notice that when k is
odd this question comes down to Warings problem for k-th powers of numbers
in Z.
Theorem 7.4.1 We have v(k) 2k1 + (k!)/2.
Proof. We introduce the dierence operator which acts on polynomials f (x)
by (f )(x) = f (x+1)f (x). Notice that decreases the degree of a polynomial
by one. Furthermore,
( ) (
) ( ) (
)
x
x+1
x
x
=
.
k
k
k
k1
Notice also that
(
)
( )
x
x
xk
k1
=
+ r(x)
+
2
k1
k!
k
where r(x) has degree k 2. Application of the operator k1 yields

( k)
x
k1
k1
x+
=
.
2
k!
Hence k1 xk = k!x + k1
k!. The expression k1 xk can be considered as the
2
sum of 2k1 terms of the form mk . Now let n N. Determine x, l Z such
that
1
k1
k! + l,
|l| k!.
n = k!x +
2
2
k1 k
Hence n = x + l. Notice that l can be written as the sum of |l| terms 1k or
1k . Since |l| (k!)/2 we obtain v(k) 2k1 + (k!)/2.
2
The bound of Theorem 7.4.1 is much larger than the actual value of v(k). If we
accept the known upper bounds for G(k) Theorem 7.4.2 yields a much better
result. However, the advantage of Theorem 7.4.1 is that it is simple and selfcontained.
80
Theorem 7.4.2 We have v(k) G(k) + 1.

Proof. Let n N. Choose y k suciently large so that n + y k can be written as
the sum of at most G(k) positive k-th powers, n + y k = xk1 + + xkG(k) . Hence
v(k) G(k) + 1.
2
Theorem 7.4.3 We have v(2) = 3 and v(3) = 4 or 5.

Proof. Theorem 7.4.1 gives us v(2) 2 + 1. Notice that also v(2) 3 because
sums of the form a2 b2 which are positive can never be 6(mod 8). Hence
v(2) = 3.
Let n N. Notice that n3 n is divisible by 6. Write n3 n = 6x. Notice also
that 6x = (x 1)3 + (x + 1)3 2x3 . Hence n = n3 (x 1)3 (x + 1)3 + 2x3
and we see that v(3) 5. Moreover, cubes are always 0, 1(mod 9), so a sum of
three cubes can never be 4(mod 9). Hence v(3) 4.
2
It is generally believed, but not proved yet, that v(3) = 4.
Another variation on Warings problem is the question, is there a number s such
that any n N can be written as the sum of s non-negative k-th powers of
rational numbers? I do not know of a simple proof for the existence of such s but
of course the positive solution to Warings problem also implies a positive answer
to our question. In fact, we can take s G(k). For the otherwise elusive case
k = 3 we have a particularly nice solution.
Theorem 7.4.4 (Riley, 1825) Any n N can be written as the sum of three
positive rational cubes in innitely many ways.
Proof. Let A = 12t(t + 1) (t + 1)3 ,
and consider the identity
B = (t + 1)3 12t(t 1),
C = 12t(t 1)
A3 + B 3 + C 3 = 72t(t + 1)6 .
Let n N. Choose u Q such that 1 < nu3 /72 < 2 and take t = nu3 /72. Then
A, B, C, t, t + 1 are all positive and we get
(
A
u(t + 1)2
)3
+
B
u(t + 1)2
)3
+
C
u(t + 1)2
Moreover, u can be choosen in innitely many ways.

)3
= n.
2
7.5. EXERCISES
7.5
81
Exercises
2
2
Exercise 7.5.1
Let r2 (n) be the number of solutions x, y Z to n = x + y .
Let R(X) = nX r2 (n). Prove that
R(X) = X + O( X),
in other words, |R(X) X|/ X is bounded (for all

X 1). (Hint: count the
number of lattice points inside the disk with radius X).
Exercise 7.5.2 Find all ways to write 425 as sum of two squares.
Exercise 7.5.3 Which numbers can be written as a dierence of two squares?
Exercise 7.5.4 (H.W.Lenstra jr.) Notice, 122 + 332 = 1233, 5882 + 23532 =
5882353. Can you nd more of such examples?
Exercise 7.5.5 Let p be a prime number such that p 1, 3(mod 8). Prove
that there exist x, y Z such that p = x2 + 2y 2 . (Hint: prove that Z[ 2] is
euclidean.)
Exercise 7.5.6 a) Show that 7 cannot be written as a sum of three squares (i.e.
g(2) > 3).
b) Show that 23 and 239 cannot be written as sum of at most 8 cubes (i.e. g(3) >
8).
c) Find N N such that N cannot be written as a sum of at most 18 fourth
powers (i.e. g(4) > 18).
Exercise 7.5.7 Let g(k) be the function from Warings problem. Prove that
g(k) 2k + [(3/2)k ] 2. (Hint: consider n = 2k [(3/2)k ] 1.)
Exercise 7.5.8 We are given the identity
6(a2 + b2 + c2 + d2 )2 = (a + b)4 + (a b)4 + (c + d)4
+(c d)4 + (a + c)4 + (a c)4
+(b + d)4 + (b d)4 + (a + d)4
+(a d)4 + (b + c)4 + (b c)4 .
Prove that g(4) 53. (Hint: write n = 6N + r and N as sum of four squares.)
Rene the argument to show that g(4) 50.
Exercise 7.5.9 Show that 5 cannot be written as the sum of 4 fourth powers of
rational numbers. (Hint: look modulo 5).
Can you nd other numbers that cannot be written as sum of 4 rational fourth
powers?
Chapter 8
Continued fractions
8.1
Introduction
Let R. The continued fraction algorithm for runs as follows.

x0 =
a0 = [x0 ],
x1 = 1/{x0 }
a1 = [x1 ],
x2 = 1/{x1 }
......
xn+1 = 1/{xn }
an = [xn ],
......
Notice that xi 1 for all i 1. The algorithm is said to terminate if {xn } = 0

for some n. Notice that
= a0 +
1
1
= a0 +
x1
a1 +
1
x2
= a0 +
1
1
a1 + a2 +
which is denoted as
= [a0 , x1 ] = [a0 , a1 , x2 ] = [a0 , a1 , a2 , . . . ] .
Theorem 8.1.1 The continued fraction algorithm terminates if and only if
Q.
Proof. Suppose we have termination, i.e. {xn } = 0 for some n. Then =
[a0 , a1 , . . . , an1 ] and we see trivially that Q.
If Q then the xi are all rational numbers, say xi = pi /qi with pi , qi N and
pi > qi for all i. Notice that qi+1 = pi [pi /qi ]qi for all i, hence q1 > q2 > q3 >
> 0. So we see that the algorithm terminates.
2
82
8.1. INTRODUCTION
83
In fact, when is rational, = p/q then the continued fraction algorithm is

nothing but the Euclidean algorithm applied to p, q.
Theorem 8.1.2 Let a0 , a1 , . . . , an R. Suppose
p2 = 0
q2 = 1
p1 = 1 p0 = a0
q1 = 0 q0 = 1
pn = an pn1 + pn2
qn = an qn1 + qn2
Then,
[a0 , a1 , . . . , an ] =
(n 0)
(n 0)
pn
.
qn
Proof. By induction on n we shall show that [a0 , . . . , an ] = (an pn1 +pn2 )/(an qn1 +
qn2 ). For n = 0 this is trivial. Now suppose n 0. Notice that
[a0 , a1 , . . . , an , an+1 ] = [a0 , a1 , . . . , an1 , an +
]
an+1
(an + 1/an+1 )pn1 + pn2
=
(an + 1/an+1 )qn1 + qn2
an+1 (an pn1 + pn2 ) + pn1
=
an+1 (an qn1 + qn2 ) + qn1
an+1 pn + pn1
=
an+1 qn + qn1
2
which completes our induction step.
From now on we shall adhere to the notations = [a0 , a1 , . . . ] , [a0 , a1 , . . . , an ] =

pn /qn for the continued fraction expansion of . We call a0 , a1 , a2 , . . . the partial
fractions of the continued fraction and the pn /qn the convergents . Why the pn /qn
are called convergents will become clear from the following theorem.
Theorem 8.1.3 Let notation be as above. Then, for all n 0,
1.
pn1 qn pn qn1 = (1)n .
2.
pn
(1)n
=
.
qn
qn (xn+1 qn + qn1 )
Proof. Part (1) is proved by induction on n, the case n = 0 being trivial.

pn pn+1 pn an+1 pn + pn1

=
pn qn+1 pn+1 qn =
qn qn+1 qn an+1 qn + qn1

pn1 pn
pn pn1
n
n+1

=
=
qn1 qn = (1) = (1)
qn qn1
Part (2) follows from = [a0 , a1 , . . . , an , xn+1 ] =
ward computation of the dierence pqnn .
xn+1 pn +pn1
xn+1 qn +qn1
and a straightfor2
84
CHAPTER 8. CONTINUED FRACTIONS
Corollary 8.1.4 For all convergents pn /qn we have

p
1
n
< 1 <

qn
qn+1 qn
an+1 qn2
We see from the previous corollary that convergents p/q of the continued fraction
of an irrational number have the property that

p

< 1
(8.1)
q
q2
In particular this means the convergent give very good rational approximations
with respect to their denominator. As an example consider
= [3, 7, 15, 1, 292, 1, 1, 1, 2, 1, . . . ].
From the theory we expect that | p/q| < 1/(292q 2 ) where p/q = [3, 7, 15, 1]
which equals 355/113. In fact,
355
= 0.000000266764
113
and 355/113 approximates up to 6 decimal places. Note that there does not
seem to be any regularity in the continued fraction of .
Here are some other examples of continued fraction expansions:
e = [2, 1, 2, 1, 1, 4, 1, 1, 6, 1, 1, 8, 1, 1, 10, . . . ]
e2 = [7, 2, 1, 1, 3, 18, 5, 1, 1, 6, 30, 8, 1, 1, 9, 42, 11, 1, 1, . . . ]
e3 = [20, 11, 1, 2, 4, 3, 1, 5, 1, 2, 16, 1, 1, 16, 2, 13, 14, 4, 6, 2, 1, 1, 2, 2, . . . ]
2 = [1, 2, 2, 2, 2, 2, 2, . . . ]
97 = [9, 1, 5, 1, 1, 1, 1, 1, 1, 5, 1, 18, 1, 5, 1, 1, 1, 1, 1, 1, 5, 1, 18, 1, . . . ]
47 = [6, 1, 5, 1, 12, 1, 5, 1, 12, 1, 5, 1, 12, 1, 5, 1, . . . ]
2
It is interesting to note the regularity in the expansions of e, e
and N , but
not in e3 . We shall reurn to the periodicity of the expansion of N in the next
section.
It also turns out that if a fraction p/q satises (8.1) then it is almost a convergent
of the continued fraction of .
Theorem 8.1.5 (Legendre) Suppose R and p, q Z, q > 0 such that

p
< 1 .

q 2q 2
Then p/q is a convergent of the continued fraction of
8.2. CONTINUED FRACTIONS FOR QUADRATIC IRRATIONALS
85
Proof. Let p0 /q0 , p1 /q1 , . . . be the convergents of the continued fraction of .

Choose n such that qn q < qn+1 . Let us assume p/q = pn /qn , otherwise we are
done. Then we have the following inequalities,

p pn
1
1
pn
p
1

+ <
+ 2
qqn
q qn
qn
q
qn qn+1 2q
q
Multiply these inequalities with qqn to obtain 1 < q2qn + qn+1
. When qn+1 2q we
nd, using qn q that 1 < 1/2 + 1/2 which is a contradiction.
So let us assume qn+1 < 2q. We now repeat our estimates with a little more
care. Suppose rst that p/q and pn /qn have the same sign. Then the
absolute value of their dierence, which equals |p/q pn /qn |, is bounded above
by max(1/2q 2 , 1/qn qn+1 ). It is bounded below by 1/qqn . Multiplication by qqn
yields 1 < max(qn /2q, q/qn+1 ) < max(1/2, 1) = 1, again a contradiction.
Now suppose that p/q and pn /qn have opposite sign. Then p/q and
pn+1 /qn+1 have the same sign. Just as above we derive 1 < max(qn+1 /2q, q/qn+2 ).
Using qn+1 < 2q we nd 1 < max(1, 1) = 1, again a contradiction.
2
Here is a very useful lemma in all that follows.

Lemma 8.1.6 Let = [a0 , a1 , a2 , . . . , am , ]. Then 1/ = [am , . . . , a2 , a1 , a0 , 1/].
Proof. This goes by induction on m. For m = 0 the lemma is clear,
= [a0 , ] = a0 +
1
1
1
1
1
= a0 +
= [a0 , ].
1/
Suppose m > 0. Then, = [a0 , . . . , am1 , am +1/]. By the induction hypothesis

we obtain
1
1
= [am1 , . . . , a1 , a0 , ]
am + 1/
Invert on both sides and add am to obtain

1
1
= [am , . . . , a2 , a1 , a0 , ]
8.2
Continued fractions for quadratic irrationals
A real, non-rational number which satises a polynomial equation of degree

2 over Q is called a quadratic irrational . Given a quadratic irrational there
exist, up to common sign change, a unique triple of integers A, B, C such that
A2 +B+C = 0 and gcd(A, B, C) = 1, AC = 0. The polynomial AX 2 +BX +C
2
is called the minimal polynomial
is called the
of . The number D = B 4AC
discrminant of . If = a + b D for some a, b Q we call a b D its conjugate
and denote it by . A quadratic irrational is called reduced if > 1 and
1 < < 0.
86
Theorem 8.2.1 Let be a quadratic irrational of discriminant D. Put x0 =

and dene recursively xn+1 = 1/(xn [xn ]). Then each xn has discriminant D
and there exists n0 such that xn is reduced for all n > n0 .
Proof. That the discrimant does not change is clear if we realise that and
m have the samen discriminant for any m Z and that and 1/ have the
same discrminant.
Denote by xn the conjugate of xn . Notice that xn > 1 for all n 1. Verify also
that if xn is reduced the same holds for xn+1 , xn+2 , . . . .
Let m be the smallest index such that [xm ] = [xm ]. If such an m would not exist,
both and have the same continued fraction expansion.
Suppose that [xm ] < [xm ]. Then notice that xm+1 = 1/(xm [xm ]) < 0. From
xm+2 = 1/(xm+1 [xm+1 ]) and xm+1 < 0 we then conclude that 1 < xm+2 < 0,
hence xm+2 is reduced.
Now suppose that [xm ] > [xm ]. Then xm+1 = 1/(xm [xm ]) < 1. Hence [xm+1 ] =
0 < [xm+1 ] and we continue as in the preceding case.
2
Theorem 8.2.2 There exist nitely many reduced quadratic irrationals of given
discriminant D.
with P, Q Z,
Proof. Let be such a quadratic irrational and write = D+P
Q
Q = 0 (we take D > 0.

From > 0> it follows that Q > 0. From > 1 > it follows
that
D + P > D P , hence P > 0. From <
0
it
follows
that
P
D < 0,
hence P < D. From > 1 we conclude
P + D > Q, hence
Q < 2 D.
Concluding, we nd that 0 < P < D and 0 < Q < 2 D, hence we have at

most nitely many possibilities.
2
Let [a0 , a1 , a2 , . . .] be the continued fraction expansion of a real number. We say
that the expansion is periodic if there exist n0 Z, N N such that an+N = an
for all n n0 . We call the expansion purely periodic if n0 = 0.
Theorem 8.2.3 Let R. Then the continued fraction expansion of is
periodic if and only if is a quadratic irrational. It is purely periodic if and only
if is reduced.
Proof. We rst prove our theorem for purely periodic expansions. Suppose
has a purely periodic continued fraction. Then there exists an r such that =
[a0 , a1 , . . . , ar , ]. Hence
pr + pr1
=
qr + qr1
This implies qr 2 +(qr1 pr )pr1 = 0. First of all we see that is a quadratic
irrational. Since its continued fraction is purely periodic we must have a0 1,
8.2. CONTINUED FRACTIONS FOR QUADRATIC IRRATIONALS
87
hence > 1. From the quadratic equation we see that = pr1 /qr . Using
either pr1 /qr = (pr1 /qr1 )(qr1 /qr ) or pr1 /qr = (pr1 /pr )(pr /qr ) we conclude
that pr1 /qr < , hence 1 < < 0. So is reduced.
Suppose conversely that is a reduced quadratic irrational. Let x0 = and
recursively xn+1 = 1/(xn [xn ]). Since x0 is reduced, all xi are reduced. Moreover
their discriminants are all the same, hence there exist only nitely many distinct
xi . So there exist r < s such that xr = xs . Notice that the value of an follows
uniquely from xn+1 by the condition that xn is reduced, namely an = [1/xn+1 ].
Hence xn = [1/xn+1 ] + 1/xn+1 . In particular it follows from xr = xs that
xr1 = xs1 , etcetera, hence x0 = xsr . So the continued fraction of x0 = is
purely periodic.
Now suppose that has a periodic continued fraction. Then there exists with
a purely periodic expansion such that = [a0 , a1 , . . . , an0 , ]. We know that
is a quadratic irrational from the above, hence the same holds for . Suppose
conversely that is a quadratic irrational. Then we know that there is a reduced
such that = [a0 , . . . , an0 , ]. Since has periodic continued fraction, the
same holds for .
2
For quadratic irrational numbers of the form N , N N not a square, we obtain

the following theorem.
Theorem 8.2.4 Let N N

and suppose N is not a square. Then N =
[a0 , a1 , . . . , ar , 2a0 ] where a0 = [ N ]. Moreover, (a1 , a2 , . . . , ar ) = (ar , . . . , a2 , a1 ).
Proof. First observe that a0 = [ N

] is the result of the rst step in the continued
fraction algorithm. Now note
that N + a0 is reduced quadratic irrational, since

N + a0 > 1 and 1 < N + a0 < 0. Hence it has a purely periodic continued
fraction of the form
N + a0 = [2a0 , a1 , a2 , . . . , ar ] = [2a0 , a1 , a2 , . . . , ar , 2a0 ].
After subtraction of a0 on
both sides we obtain N = [a
0 , a1 , . . . , ar , 2a0 ], as
asserted. Notice also that N + a0 = [2a0 , a1 , a2 , . . .
, ar , N + a0 ]. Substract
2a0 on both sides to nd N a0 = [0, a1 , a2 , . . . , ar , N + a0 ]. Hence
1
= [a1 , a2 , . . . , ar , N + a0 ].
N a0
Application of Lemma 8.1.6 yields
1
= [ar , . . . , a2 , a1 , a0 N ]
N + a0
This algebraic identity remains true if we replace N by N ,
1

= [ar , . . . , a2 , a1 , a0 + N ].
N + a0
88
Invert both sider and add 2a0 to obtain
N + a0 = [2a0 , ar , . . . , a2 , a 1, N + a0 ].
So we see that the continued fraction of N +a0 is also given by [2a0 , ar , . . . , a2 , a1 ].

Hence (a1 , a2 , . . . , ar ) = (ar , . . . , a2 , a1 ).
2
8.3
Pells equation
Suppose N N is not a square and consider the diophantine equation

x2 N y 2 = 1
in the unknowns x, y Z0 . Although problems related to his equation have
been around since antiquity, the rst general method for solving it was given
by W.Brouncker in 1657. He was able to use his method to obtain the smallest
solution
(x, y) = (32188120829134849, 1819380158564160)
to x2 313y 2 = 1! Brounckers method was described in Walliss book on algebra and number theory. Euler mistakenly assumed from Walliss book that the
method was due to John Pell, another English mathematician. Very soon Pells
name stuck to this equation. For several values of N we list the solution with
minimal x,
32 2 22 = 1
6492 13 1802 = 1
17663190492 61 2261539802 = 1
Looking at these examples one observes that it is quite a miracle that any nontrivial solution for x2 61y 2 = 1 exists. Nevertheless, using continued fractions
it is possible to show that there always exists a non-trivial solution.
Proposition 8.3.1 Let N N and suppose N is not a square. Then there exist
x, y N such that x2 N y 2 = 1.
Proof. For N = 2, 3, 5, 6 our theorem is true since we have 32 2 22 = 1,
22 3 12 = 1, 92 5 42 = 1, 52 6 22 = 1.So we can assume that N 7.
Consider the continued fraction expansion of N given by
N = [a0 , a1 , . . . , ar , 2a0 ]
say. Let p/q = [a0 , a1 , . . . , ar ]. Then, from our elementary estimates we nd that

p
N < 1 .
2a0 q 2
q
8.3. PELLS EQUATION
89
Multiply
on
both
sides
by
|p/q
+
N
|
and
use
the
fact
that
|p/q
+
N|
(2 N + 1). We nd,
2

p

N < 2 N + 1.
q2

2a0 q 2
Multiply on both sides by q 2 to nd |p2 N q 2 | < (2 N + 1)/2[ N ]. When

N 7 we have
2 N +1
2 N +1
<
<2
2[ N ]
2( N 1)
Hence |p2 N q 2 | < 2. So we have either p2 N q 2 = 1 or p2 N q 2 = 1. (why cant
we have p2 N q 2 = 0?). In case p2 N q 2 = 1 we nd x = p, y = q as solution.
In case p2 N q 2 = 1 we notice that (p2 + N q 2 )2 N (2pq)2 = (p2 N q 2 )2 = 1.
Hence we have the solution x = p2 + N q 2 , y = 2pq.
2
Now that we established the existence of non-trivial solutions to Pells equation
we would like to have the full set. An important remark to this end is the following
trick which
by an example. Notice that 32 2 22 = 1 is equivalent
we illustrate
to (3 + 2 2)(3 2 2) = 1. Take the square on both sides and use the fact that
(3 2 2)2 = 17 12 2
Hence (17 + 12 2)(17

172 2 122 = 1. We can also take
12 2) which implies
the cube of (3 + 2 2) to obtain 99 + 70 2. We then nd 992 2 702 = 1. So,

given one solution of Pells equation we can construct innitely many! If we start
with the smallest positive solution we get all solutions in this way, as shown in
the following theorem.
Theorem 8.3.2 Choose the solution of Pells equation with x + y N > 1 and
minimal. Call it (p, q). Then,
x, y N of Pells equation there
to any solution
exists n N such that x + y N = (p + q N )n .
2
2
Proof.
Notice that if u, v Z satisfy
1u N v = 1 and u + v N 1, then
u v N , being
equal to (u + v N )
lies between 0 and 1. Addition of the
inequalities u + v N 1 and 1 u v N
> 0 implies u 0. Substraction of
these inequalities yields v > 0. We call u + v N the size of the solution
u, v.
Now
let x, y N be any solutionof Pells equation. Notice that (x + y N )(p
q N ) = (px qyN ) + (py
qx) N . Let
u = px qyN, v = py qx and we have
2
2
u N v = 1 and u + v N = (x + y N )/(p + q N ). Observe that
x+y N
< (x + y N )/2,
1
p+q N
hence 1 <= u + v N < (x + y N )/2. So we have found anew solution with

positive coordinates and size bounded by half the size of x + y N . By repeatedly
90
performing
this operation we obtain a solution whose size is less than the size of
p + q N . By the minimality of p, q this implies that this last solution

should
be 1, 0.
the number of steps is n we thus nd that x + y N =
Supposing
n
2
(p + q N ) .
In the existence
proof for solutions to Pells equation we have used the continued
fraction of N . It turns out that we can use this algorithm to nd the smallest
solution and also all solutions of other equations of the form x2 N y 2 = k for
small k.
Theorem 8.3.3 Suppose we have x, y N suchthat |x2 N y 2 | N . Then

x/y is a convergent to the continued fraction of N .
Proof. Let M = [ N ]. Since x2 N y 2 is integral the inequality |x2 N y 2 | < N

implies |x2 N y 2 | M . We rst show that x M y. If x < M y we would have
the following sequence of inequalities,
x2 N y 2 < x2 M 2 y 2 = (x yM )(x + yM ) < M
contradicting |x2 N y 2 | M . So we have x M y.
Notice that |x2 N y 2 | M implies
|x y N |
M
M
M
1
<
=
x + yM
2yM
2y
x+y N
Divide by y on both sides and use Theorem 8.1.5 to conclude that x/y is a
convergent.
2
8.4
Archimedess Cattle Problem
The following story has been taken from Albert H.Beiler, Recreations in the
Theory of Numbers, Dover. It deals with a problem which is attributed to
Archimedes. In the form of an epigram one is asked for the number of oxen
of the sun. The oxen go in four colors, white, black, spotted and yellow. Let
W, X, Y, Z be the number of bulls of color white, balck, spotted, yellow respectively and x, y, z, w the number of cows. It is asked that
9
X = 20
Y +Z
W = 65 X + Z
7
Y = 13
W
+
Z
w
=
(X
+ x)
42
12
11
9
x = 20 (Y + y) y = 30 (Z + z)
13
z = 42
(W + w)
and in addition W + X should be a square and Y + Z a triangular number (i.e.

of the form n(n + 1)/2. The rst seven conditions are easy and one veries that
8.5. CORNACCHIAS ALGORITHM
91
there exists k N such that

W = 10366482k
X = 7460514k
Y = 7358060k
Z = 4149387k
w = 7206360k
x = 4893246k
y = 3515820k
z = 5439213k
The condition W + X a square implies that 17826996k must be a square. Hence

k = 4456749t2 for some integer t. Finally the condition Y + Z = n(n + 1)/2
implies
51285802909803t2 = n(n + 1)/2
Multiply by 8 on both sides and add 1 to obtain
4n2 + 4n + 1 = (2n + 1)2 = 410286423278424t2 + 1
= 4729494(9314t)2 + 1
In other words we are looking for the solutions of the pellian equation
u2 4729494v 2 = 1
where v must be divisible by 9314. Finding the smallest solution has only been
possible by the use of a computer and the smallest solution value of u turns out
to have 206554 decimal digits.
8.5
Cornacchias algorithm
Closely related to the previous section is the problem to write a prime number in
the form x2 + dy 2 , where d N is given. It is known for example, that p is a sum
of two squares if and only if p 1(mod 4). The question is how to nd the two
squares if p 1(mod 4). One approach would be as follows. Determine z N
such that z 2 d(mod p) by the method of the previous section. Then consider
the lattice L generated by the vectors (z, 1) and (p, 0). Moreover, for every (a, b)
L we have a2 + db2 0(mod p). Conversely, suppose that a2 + db2 = p is solvable
in integers a, b. Then it follows from a2 db2 (mod p) that a zb(mod p).
Hence either (a, b) or (a, b) is in the lattice L. Suppose (a, b) L. If we take the
norm x2 + dy 2 on L then (a, b) is a shortest vector in L and we can use reduction
in dimension 2 to nd this shortest vector.
However, the reduction algorithm with this particular application can be reformulated in an even simpler way.
Theorem 8.5.1 (G.Cornacchia,1908) Let d N and p a given odd prime.
Suppose a2 + db2 = p is solvable in the positive integers a, b. If d = 1 we assume
that a > b. Let x0 N be such that x20 d(mod p) and x0 < p/2. Apply the
following algorithm recursively, x1 = p[p/x0 ]x0 , x2 = x0 [x0 /x1 ]x1 , . . . , xi+1 =
xi1 [xi1 /xi ]xi , . . . . Choose i such that xi b0 if d = 1.

Note that (a0 , b0 ), (db0 , a0 ) span a lattice L of determinant p. Note also that
if (x, y) = (a0 , b0 ) + (db0 , a0 ), then x2 + dy 2 = (2 + d2 )p. In particular,
(x, y) L x2 + dy 2 0(mod p).
Let (a1 , b1 ) = (db0 , a0 ) + [a0 /b0 ](a0 , b0 ) and dene recursively
(ai+1 , bi+1 ) = (ai1 , bi1 ) + [|bi1 /bi |](ai , bi ),
i 1.
Notice that a0 < a1 < a2 < and that |b0 | > |b1 | > |b2 | > and that
(1)i bi > 0 as long as bi = 0. Notice also that, given ai+1 and ai , one can recover
ai1 by the conditions ai1 ai+1 (mod ai ) and 0 < ai1 < ai .
Suppose that bk = 0, which will indeed happen for some k. The algorithm
then
terminates.
Since each pair (ai , bi ), (ai1 , bi1 ) forms a basis of L, we have

ai1 bi1
= p. In particular when i = k we obtain ak bk1 = p. Hence

ai
bi
ak divides p. On the other hand, 0 a2k + db2k a2k (mod p), hence p divides
ak . So we get ak = p and bk1 = 1. Hence, from ak1 + db2k1 0(mod p)
we get a2k1 d(mod p). Starting with the values ak , ak1 we can compute
ak2 , ak3 , . . . recursively. If ak1 = x0 this will be exactly the recursion procedure
of our theorem. If ak1 = px0 , we note that ak2 = x0 , ak3 = px0 (mod x0 ) =
p(mod x0 ) and the recursion again coincides with the recursion of our theorem.
We have trivially a0 < p. If we can show that a1 > p we know exactly at

which point in the euclidean algorithm for p, x0 we have
hit upon the desired
a0 . To show that a1 > p notice that |b1 | < b0 | < p/d. Moreover, a short
computation shows that a1 + db21 = ([a0 /b0 ]2 + d)p. Using b21 < p/d this implies
a21 > ([a0 /b0 ]2 + d 1)p. When d > 1 we are done. If d = 1 we know that a0 > b0 ,
hence [a0 /b0 ] 1 and we are also done.
2
In the case d = 1 the algorithm has some nice side properties. Suppose that
x0 1(mod p) and 0 < x0 < p/2. Then the continued fraction algorithm
of p/x0 is symmetric of even length. Moreover, let a and b be the rst two
remainders less than p in the euclidean algorithm for p/x0 . Then a2 + b2 = p.

All this, and more, is shown in the exercises below.
First we treat the example of writing the prime p = 1020 + 129 as sum of two
squares. Note that m := (p 1)/27 is odd. We rst determine a generator g of
the group of elements modulo p whose order divides 27 . Then, noting that any
solution x of x2 1(mod p) has order 4, we simply take x g 32( as) solution.
( )
The smallest quadratic nonresidue modulo p is 7. We easily check p7 = p7 =
(5)
= 1. Then a generator g can be found by setting g 7m (mod p), and a
7
solution of x2 1(mod p) by x 732m 44237909966037281987(mod p). The
continued fraction of p/x equals
[2, 3, 1, 5, 5, 167, 3, 14, 69, 33, 2, 2, 33, 69, 14, 3, 167, 5, 5, 1, 3, 2]
8.6. EXERCISES
93
and Cornacchias algorithm yields after 11 steps,

p = 89708788732 + 44185215002 .
8.6
Exercises
A small point to be recalled in the following exercises is that the continued fraction
of a rational number is not unique. If we would have p/q = [a0 , . . . , ar , 1] for
example, we can rewrite it as p/q = [a0 , . . . , ar +1]. This would be the result of the
continued fraction algorithm. In the exercises we shall use the word normalised
continued fraction if we want the last partial quotient larger than 1.
Exercise 8.6.1 Let p, q N be relatively prime and such that q < p. Denote the
continued fraction of p/q by [a0 , . . . , ar ]. Let q be such that qq (1)r (mod p).
Using Lemma 8.1.6, show that p/q = [ar , . . . , a0 ].
Exercise 8.6.2 Suppose the rational number p/q has a symmetric continued
fraction, i.e. p/q = [a0 , a1 , . . . , am ] with (a0 , a1 , . . . , am ) = (am , . . . , a1 , a0 ). Using
Lemma 8.1.6, show that q 2 (1)m (mod p).
Exercise 8.6.3 Let p, q N such that gcd(p, q) = 1, q < p/2 and q 2 (mod p).
Show that the normalised continued fraction expansion of p/q is symmetric.
Exercise 8.6.4 Let b/a = [a0 , a1 , . . . , ar ] where ai N for all i and gcd(a, b) = 1.
Let p/q = [ar , ar1 , . . . , a0 , a0 , . . . , ar ]. Using Lemma 8.1.6 show that p = a2 + b2 .
Chapter 9
Diophantine equations
9.1
General remarks
Let F (x1 , . . . , xr ) Z[x1 , . . . , xr ]. An equation of the form

F (x1 , . . . , xr ) = 0
(9.1)
in the unknowns x1 , . . . , xr in Z or Q is called a diophantine equation.

Examples are:
x7 + y 7 = z 7 ,
x2 67y 2 = 1,
y 2 = x3 2.
The behaviour of the solution sets seems to be very erratic and depends very
strongly on small variations of the coecients. For example x3 +y 3 = z 3 and x3 +
y 3 = 4z 3 are known to have no integral solutions with xyz = 0 whereas x3 + y 3 =
13z 3 and x3 + y 3 = 22z 3 have innitely many with xyz = 0 and (x, y, z) = 1 (the
smallest solutions being (x, y, z) = (2, 7, 3), (25469, 17299, 9954) respectively).
A famous problem on diophantine equations was Hilberts tenth problem : Is
there a computer program, using unlimited memory, with which one can decide
whether any equation of the form (9.1) has a solution or not. We use the term
computer program here to avoid having to explain the denition of algorithm.
The answer to this question, given in 1970 by Matijasevich is no. The proof is
based on a combination of logic and number theory and unfortunately falls outside
the scope of these notes. The result of Matijasevich suggests that any diophantine
equation has its own peculiarities. On the one hand it makes the solution of
diophantine equations harder, but on the other hand also more interesting because
of the variety of approaches which are now required. In the following sections we
shall discuss a few classes of diophantine equations.
9.2
Pythagorean triplets
One of the ancient diophantine equations is the following.

94
9.2. PYTHAGOREAN TRIPLETS
95
Definition 9.2.1 A triplet a, b, c N is called Pythagorean if

gcd(a, b, c) = 1
and
a2 + b2 = c2 .
Notice that in a Pythagorean triplet a and b cannot be both odd. For then
we would have a2 + b2 1 + 1 2(mod 4) but c2 , being a square, cannot be
2(mod 4).
Theorem 9.2.2 Let a, b, c N and suppose that b is even. Then a, b, c is a
Pythagorean triplet if and only if r, s N : r > s, r s(mod 2), (r, s) =
1, a = r2 s2 , b = 2rs, c = r2 + s2 .
Proof. Suppose we have r, s with the given properties. Clearly a, b, c satisfy
a2 + b2 = c2 . Notice also that (a, c) divides (c a, c + a) = (2s2 , 2r2 ) = 2. But
since r s(mod 2), a and c are odd and so (a, c) = 1. Hence (a, b, c) = 1.
Suppose now that a, b, c is a pythagorean triplet. Write b2 = (c a)(c + a). Since
a, c are both odd this implies (b/2)2 = ((a + c)/2)((c a)/2). From (a, b, c) = 1
and a2 + b2 = c2 it follows that (a, c) = 1. Hence also ((c + a)/2, (c a)/2) = 1.
Since the product of these numbers equals (b/2)2 each of them is a square, say
(c + a)/2 = r2 and (c a)/2 = s2 . Hence c = r2 + s2 , a = r2 s2 . Moreover,
(r, s) = 1 and r2 (c + a)/2 s2 + a s2 + 1(mod 2) r2 s2 (mod 2) r
s(mod 2).
2
When we rewrite the equation a2 + b2 = c2 as (a/c)2 + (b/c)2 = 1 we see that
nding Pythagorean triplets is equivalent to nding rational numbers p, q such
that p2 + q 2 = 1, in other words, nding rational points on the unit circle.
Geometrically, the solution to this problem runs as follows. For any point (p, q)
Q2 we draw the line between (p, q) and (1, 0) which is given by Y = t(1 X),
where t = q/(1 p). Conversely, any line through (1, 0) is given by Y = t(1 X).
2 1
The second point of intersection with the unit circle is given by tt2 +1
, t22t+1 . Thus
we can conclude: there exists a bijection between the sets
{t Q} and {x, y Q| x2 + y 2 = 1, (x, y) = (1, 0)}
given by
y
t=
,
1x
(
(x, y) =
t2 1 2t
,
t2 + 1 t2 + 1
)
.
Using Theorem 9.2.2 it is very simple to nd all , , Q such that 2 + 2 = 2 .

They are all of the form
( 2
)2 (
)2 ( 2
)2
r + s2
r s2
2rs
=
+
, r, s, M Z.
M
M
M
A number n N is called congruent if n is the surface area of a right-angled
triangle with rational sides. In other words
n is congruent r, s, M Z such that n =
1 2rs r2 s2
.
2M M
96
CHAPTER 9. DIOPHANTINE EQUATIONS
The latter equation is equivalent to M 2 n = rs(r2 s2 ). It is a classical problem

to characterise congruent numbers. N.Koblitz used this topic in his book Introduction to Elliptic Curves and Modular Forms as a leading motive. The smallest
congruent number is 5. A notorious congruent number is n = 157. The simplest
triangle corresponding to it has right angle sides
6803298487826435051217540
411340519227716149383203
411340519227716149383203
21666555693714761309610
and hypothenusa
224403517704336969924557513090674863160948472041
8912332268928859588025535178967163570016480830
(D.Zagier)
As an interesting curiosity we mention two results on congruent numbers

Theorem 9.2.3 (Birch, 1975) When n is prime and equal 5 or 7 modulo 8
then n is congruent. When n is twice a prime of the form 1(mod 4) then n is
also congruent.
Theorem 9.2.4 (Tunnell, 1983) Suppose n is a congruent number then the
number of integral solutions to 2x2 + y 2 + 8z 2 = n equals twice the number of
solutions to 2x2 + y 2 + 32z 2 = n.
It is generally expected that the converse of Tunnells theorem also holds.
9.3
Fermats equation
After reading about pythagorean triples it seems natural to ask the following
question. Let n N and n > 2. Does the equation
xn + y n = z n
(9.2)
have any solutions in x, y, z N? Fermat believed the answer to be no and

claimed to have a remarkable proof. Unfortunately the margin of the book in
which he made this claim was too narrow to write this proof down. In June
1993 the English mathematician Andrew Wiles came quite close and for some
time it was believed that he did have a proof. However, his 200 page manuscript
of highly advanced mathematics turned out to have a gap and it took about a
year of suspense before this gap was repaired with the help of R.Taylor, a former
student of Wiles. This happened in October 1994. Wiless work not only resolves
Fermats last problem, it is also a major advance in the theory of elliptic curves,
in particular the Shimura-Taniyama-Weil conjecture.
Before Wiless discovery the equation (9.2) had been solved for certain special
values of n. For example, Fermat did prove the following theorem.
Theorem 9.3.1 The equation x4 + y 4 = z 2 has no solution x, y, z N.
9.3. FERMATS EQUATION
97
As a consequence we see that (9.2) with n = 4 has no solutions.

Proof. Suppose there exists a solution. Let x0 , y0 , z0 be a solution with minimal
z0 . We may assume that (x0 , y0 ) = 1 and that y0 is even. We shall repeatedly
use Theorem 9.2.2. From x40 + y04 = z02 follows,
r, s Z : (r, s) = 1, x20 = r2 s2 , y02 = 2rs, z0 = r2 + s2 .
From x20 + s2 = r2 , x0 odd and (r, s) = 1 follows,
, Z : (, ) = 1, x0 = 2 2 , s = 2, r = 2 + 2 .
Together with y02 = 2rs this yields (y0 /2)2 = (2 + 2 ). Since the factors
, , 2 + 2 are pairwise relatively prime, and their product is a square, we get
u, v, w Z : = u2 , = v 2 , 2 + 2 = w2 .
After
elimination
of , we get w2 = u4 + v 4 . A simple check shows |w| =
2
2
+ = r < y0 < z0 , contradicting the minimality of z0 . Hence there can
be no solutions.
2
The principle to construct a smaller solution out of a given (hypothetical) solution
is known as Fermats descending induction or descent . This principle, in disguised
form with cohomology groups and all, is still often used for many diophantine
equations.
The case n = 3 was settled by Euler (1753), Dirichlet dealt with the case n = 5
in 1820 and Lame proved Fermats conjecture for n = 7 in 1839. Notice that
the case n = 6 follows from n = 3 because x6 + y 6 = z 6 can be rewritten as
(x2 )3 + (y 2 )3 = (z 2 )3 . In general, since any number larger than 2 is divible either
by 4 or by an odd prime, it suces to prove Fermats conjecture for n = 4, which
we have already done, and for n prime. The methods of solution all follow the
same pattern. Let p be an odd prime and put = e2i/p . Then xp + y p = z p can
be rewritten as
(x + y)(x + y) (x + p1 y) = z p .
The left hand side of the equation has been factored into linear factors at the
price of introducing numbers from Z[]. The right hand side of the equation is a
p-th power and the principle of the proof is now to show that the linear factors
on the left are essentially p-th powers in Z[]. To reach such a conclusion we
would need the property that factorisation into irreducible elements is unique in
Z[]. Assuming this one would be able to conclude a proof of Fermats conjecture,
although it is still not easy. Unfortunately there is one more complicating factor,
prime factorisation in Z[] need not be unique. Finding a way around this problem
has been one of the major stimuli to the development of algebraic number theory.
In 1847 E.Kummer proved the following remarkable theorem.
98
Theorem 9.3.2 (Kummer) Denote by B0 , B1 , B2 . . . the sequence of Bernoulli

numbers. If the odd prime number p does not divide the numerators of B2 , B4 , B6 , . . . , Bp3
then xp + y p = z p has no solution in positive integers.
Recall that the Bernoulli numbers B0 , B1 , B2 , . . . are given by the Taylor series
Bn
x
=
xn .
ex 1 n=0 n!
It is not hard to see that Bn = 0 when n is odd and larger than 1. A small list
of values,
B2 = 1/6
B4 = 1/30
B6 = 1/42
B8 = 1/30
B10 = 5/66
B12 = 691/2730
B14 = 7/6
B16 = 3617/510
B18 = 43867/798
B20 = 174611/330
As an amusing aside we mention that the numerator of Bk with k p 3, k
even, is divisible by p if and only if 1k + 2k + + (p 1)k is divisible by p2 .
Using the computer and further renements of Kummers theorem one had been
able to verify Fermats conjecture for 2 < n < 4 106 (Buhler, Crandall, Sompolski) around 1990. For more details about the history and proof of Kummers
theory we refer to the books of P.Ribenboim (13 Lectures on Fermats last theorem, Springer Verlag 1979) and H.M.Edwards (Fermats last theorem, Springer
Verlag 1977). Of course these books were pre-Wiles. For an introduction for a
general audience to the techniques entering Wiless proof I highly recommend Simon Singhs book Fermats Enigma: The epic quest to solve the worlds greatest
matehmatical problem (1998). It reads like a novel.
As a generalisation of Fermats conjecture Euler conjectured that for any k
N there are no positive integers x1 , x2 , . . . , xk such that xk1 + + xkk1 = xkk .
However, this was disproved by a counterexample of Lander and Parkin (1967)
reading 1445 = 275 + 845 + 1105 + 1335 . Only in 2004 a second example was
discovered by J.Frye:
555 + 31835 + 289695 + 852825 = 853595 .
In 1988 N.Elkies found spectacular counterexamples in the case k = 4, the smallest of which reads 958004 + 2175194 + 4145604 = 4224814 . He also showed that
there exist innitely many of such examples with k = 4.
9.4
Mordells equation
Let k Z. The equation
y 2 = x3 k
(9.3)
9.4. MORDELLS EQUATION
99
in x, y Z is known as Mordells equation . This equation has been the subject of

many investigation by many people. In fact, a whole book has been written about
it (London, Finkelnstein: Mordells equation x3 y 2 = k). The main theorem is,
Theorem 9.4.1 (Mordell) The eqation (9.3) has nitely many solutions.
The proof uses algebraic number theory and is beyond the scope of these notes.
Although Mordells theorem is a niteness theorem, one cannot deduce an algorithm from it to actually determine the solutions of any given equation. Bounds,
which in principle give an eective solution became available around 1968 by
4
A.Baker who showed that log |x| c |k|10 and slightly improved by H.Stark
log |x| C |k|1+ for every > 0. The constants c, c can be computed explicitly.
It turns out that the solution set depends in a very erratic way on the value of
k. For example a short computer search reveals the solutions
32
42
52
92
232
2822
3752
3786612
=
=
=
=
=
=
=
=
(2)3 + 17
(1)3 + 17
23 + 17
43 + 17
83 + 17
433 + 17
523 + 17
52343 + 17
It is a highly non-trivial task to show that this the complete solution set of
y 2 = x3 + 17. Two examples which are easier to deal with are given in the
following theorem.
Theorem 9.4.2 The equation y 2 = x3 + 7 has no solutions in x, y Z. The
only integral solutions to the equation y 2 = x3 2 are (x, y) = (3, 5) (Fermat).
Proof. First we deal with y 2 = x3 + 7. Note that x is odd, because x even would
imply that y 2 7(mod 8), which is impossible. Now notice that
y 2 + 1 = x3 + 8 = (x + 2)(x2 2x + 4)
Notice also that for any x, x2 2x+4 = (x1)2 +3 3(mod 4). Hence x2 2x+4
always contains a prime divisor p which is 3(mod 4). So we get y 2 +1 0(mod p)
which is impossible because of p 3(mod 4).
To deal with y 2 = x3 2 we use arithmetic in the euclidean ring R = Z[ 2].

Notice rst of all that x is odd. If x were even, then x3 2 2(mod 4), so x3 2
cannot be a square. From y 2 + 2 = x3 follows the factorisation
(y + 2)(y 2) = x3
100
The gcd of y + 2 and y 2 2 divides their dierence which is 2 2. So the

gcd is either 1 or divisible by 2. Since y is odd
holds. Thus
the rst possibility
3
we nd that there
+ 2 = (a + b 2) . Computing
exist3 a, b 2 Z such2 that 2y
thecube, y + 2 = a 6ab + b(3a 2b ) 2. Comparison of the coecients
of 2 on both sides yields 1 = b(3a2 2b2 ). Hence b = 1 and 3a2 2b2 = 1.
So we nd a = 1 and b = 1. Hence x = a2 +2b2 = 3. The values of y follow. 2
An interesting dierence between the equations y 2 = x3 +7 and y 2 = x3 2 is that
the second equation has innitely many rational solutions. This can be seen by
the so-called chord and tangent method. In the point (3, 5) of the algebraic curve
y 2 = x3 2 we draw the tangent to the curve. It is given by y2 = (27/10)(x3).
Now intersect this line with the curve y 2 = x3 2. Elimination of y yields
x3
729 2 837
1161
x +
x
=0
100
50
100
Because of our tangent construction we already know that this equation has a
double root in x = 3. So the third root must also be a rational number. And
indeed we nd
129
(x 3)2 (x
)=0
100
. So the x coordinate of the third intersection point of the tangent with the curve
equals 129/100. The corresponding y coordinate is 383/1000. Indeed we check
that (x, y) = (129/100, 383/1000) is a rational solution of y 2 = x3 2. Repetition
of this procedure provides us with an innite set of rational solutions. In fact
it turns out that the rational points on y 2 = x3 2 together with the point at
innity have a group structure known as the Mordell-Weil group. This is the
beginning of a fascinating subject of rational points on elliptic curves. Excellent
introductions can be found in Silverman and Tate: Rational points on elliptic
curves.
By checking results for a large number of k M.Hall made the following conjecture
Conjecture 9.4.3 (Hall) There exists a constant C > 0 such that |x3 y 2 | >
Cx1/2 for any x, y N with x3 y 2 = 0.
It is also known that there exist innitely many positive integers x, y such that
1/2
0 < |x3 y 2 | < x (Danilov, 1982) so in this sense Halls conjecture is the
sharpest possible.
9.5
The abc-conjecture
In 1986 Masser and Oesterle formulated a striking conjecture, the truth of which
has far reaching consequences for diophantine equations. For any a Z we let
N (a) (the conductor or radical of a) denote the product of all distinct primes of
a.
9.5. THE ABC-CONJECTURE
101
Conjecture 9.5.1 (abc conjecture) Let > 0. Then there exists c() > 0
such that for any triple of non-zero numbers a, b, c Z satisfying a + b + c = 0
and gcd(a, b, c) = 1 we have
max(|a|, |b|, |c|) < c()N (abc)1+ .
To get a feeling for what this conjecture says it is best to consider a number of
consequences.
Consequence 1. Let p, q, r be xed numbers larger than 1 and (p, q, r) = 1.
Then
px + q y = r z
has only nitely many solutions x, y, z Z0 . Application of the conjecture
shows that
rz < c()N (px q y rz )1+ c()N (pqr)1+ .
Hence rz is a bounded number and so are px , q y . In particular x, y, z are bounded.
By other methods it is indeed possible to show that px + q y = rz has nitely
many solutions.
Consequence 2. Fermats conjecture is true for suciently large n. Apply the
abc conjecture to xn + y n = z n with x, y, z N to obtain
z n < c()N (xn y n z n )1+ c()N (xyz)1+ c()z 3(1+) .
Hence, assuming z 2,
2n3(1+) z n3(1+) c()
and this implies n log c()/ log 2 + 3(1 + ).
Consequence 3 Let p, q, r Z2 . Suppose
xp + y q = z r
has innitely many solutions x, y, z N with gcd(x, y, z) = 1. Then
1 1 1
+ + 1.
p q r
Application of the abc conjecture yields
z r c()N (xp y q z r )1+
c()(xyz)1+
c()(z r/p z r/q z)1+ .
Taking z this implies r (1 + r/p + r/q)(1 + ) for any > 0. Hence
r 1 + r/p + r/q and our assertion follows.
Considering the potential consequences of Conjecture 9.5.1 it is likely to be very
dicult to prove. In fact, any weaker version with 1 + replaced by another
number would already be spectacular! The best that can be done by present
day methods (1994) is max(|a|, |b|, |c|) < exp(N (abc)15 ) where is some (large)
constant.
102
9.6
The equation xp + y q = z r
Motivated by consequence 3 of the previous section and the Fermat conjecture

we might look at the equation
xp + y q = z r
gcd(x, y, z) = 1,
xyz = 0
where p, q, r are given integers > 1. From the abc-conjecture we expect this
equation to have nitely many solutions when 1/p + 1/q + 1/r < 1. It was a very
pleasant surprise when Darmon and Granville actually proved this statement in
1993. It turns out to be a consequence of Mordells conjecture, now known as
Faltingss theorem (1983), which we will describe in the next section. Until now
(1997) the only known solutions to this type of equation are
1k + 23 = 32 (k > 5)
132 + 73 = 29
27 + 173 = 712
25 + 72 = 34
35 + 114 = 1222
177 + 762713 = 210639282
14143 + 22134592 = 657
338 + 15490342 = 156133
438 + 962223 = 300429072
92623 + 153122832 = 1137 .
Notice that in each case there occurs an exponent 2. This leads us to the following
unsolved question, which can be seen as a generalisation of Fermats conjecture.
Question 9.6.1 Suppose p, q, r are integers 3. Do there exist solutions to
xp + y q = z r
gcd(x, y, z) = 1,
xyz = 0 ?
When 1/p + 1/q + 1/r = 1 we can show that the set {p, q, r} equals one of the
sets {3, 3, 3}, {2, 4, 4} or {2, 3, 6}. The case p = q = r = 3 is known to have no
solution since Euler, The case p = q = 4, r = 2 was proved in these notes. In the
exercises we show that x4 + y 2 = z 4 has no non-trivial solutions. As for the case
{2, 3, 6} we only have the solutions 23 + (1)6 = (3)2 , but this is not easy to
prove.
When 1/p + 1/q + 1/r > 1 we can show that the set {p, q, r} equals one of the
sets {2, 2, k} (k 2), {2, 3, 3}, {2, 3, 4} or {2, 3, 5}. The case p = q = r = 2
corresponds of course to pythagorean triplets. We might wonder if the other
9.6. THE EQUATION X P + Y Q = Z R
103
cases allow parametric solutions as well. In 1993 Don Zagier, amused by this
question, showed that the following parametrisation yield all integral solutions in
the cases {2, 3, 3} and {2, 3, 4}.
The equation x3 + y 3 = z 2 .
x = s4 + 6s2 t2 3t4
y = s4 + 6s2 t2 + 3t4
z = 6st(s4 + 3t4 )
x = (1/4)(s4 + 6s2 t2 3t4 )
y = (1/4)(s4 + 6s2 t2 + 3t4 )
z = (3/4)st(s4 + 3t4 )
x = s4 + 8st3
y = 4s3 t + 4t4
z = s6 20s3 t3 8t6
x = (s2 3t2 )(s4 + 18s2 t2 + 9t4 )
y = (s4 + 2s2 t2 + 9t4 )(s4 30s2 t2 + 9t4 )
z = 4st(s2 + 3t2 )(s4 6s2 t2 + 81t4 )(3s4 2s2 t2 + 3t4 )
x = 6st(s4 + 12t4 )
y = s8 168s4 t4 + 144t8
z = (s4 12t4 )(s8 + 408s4 t4 + 144t8 )
x = 6st(3s4 + 4t4 )
y = 9s8 168s4 t4 + 16t8
z = (3s4 4t4 )(9s8 + 408s4 t4 + 16t8 )
x = s6 + 40s3 t3 32t6
104

y = 8st(s3 16t3 )(s3 + 2t3 )
z = s12 176s9 t3 5632s3 t9 1024t12
x = 5s6 + 6s5 t + 15s4 t2 60s3 t3 + 45s2 t4 18st5 + 9t6
y = 6s8 56ts7 + 112t2 s6 168t3 s5 + 252t4 s4 168t5 s3 + 72t7 s 18t8
z = 29s12 +156ts11 726t2 s10 +2420t3 s9 4059t4 s8 +3960t5 s7 2772t6 s6 +
2376t7 s5 3267t8 s4 + 3564t9 s3 1782t10 s2 + 324t11 s + 27t12
x = s6 + 6s5 t 15s4 t2 + 20s3 t3 + 15s2 t4 + 30st5 17t6
y = 2s8 8ts7 56t3 s5 28t4 s4 + 168t5 s3 112t6 s2 + 88t7 s + 42t8
z = 3s12 +12ts11 66t2 s10 44t3 s9 +99t4 s8 +792t5 s7 924t6 s6 +2376t7 s5
1485t8 s4 1188t9 s3 + 2046t10 s2 156t11 s + 397t12
x = (s2 + 3t2 )(s4 18s2 t2 + 9t4 )
y = 4st(s2 3t2 )(s4 + 6s2 t2 + 81t4 )(3s4 + 2s2 t2 + 3t4 )
z = (s4 2s2 t2 + 9t4 )(s4 + 30s2 t2 + 9t4 )
x = 6st(s4 12t4 )
y = (s4 + 12t4 )(s8 408s4 t4 + 144t8 )
z = s8 + 168s4 t4 + 144t8
x = 6st(3s4 4t4 )
y = (3s4 + 4t4 )(9s8 408s4 t4 + 16t8 )
z = 9s8 + 168s4 t4 + 16t8
x = (3/2)st(s4 3t4 )
y = (1/8)(s4 + 3t4 )(s8 102s4 t4 + 9t8 )
z = (1/4)(s8 + 42s4 t4 + 9t8 )
9.7. MORDELLS CONJECTURE
105
We shall deal with the case {2, 2, k} in the exercises. The only equation that
remains is x5 + y 3 = z 2 . In 1995 it was proved by F.Beukers that again a nite
numbers of parametrised solutions suce to give the complete solution set. In
2001 Johnny Edwards managed to produce the full list of these parametrisations.
Here is the x-coordinate of one such parametrisation,
x = 185s12 144s11 t 2046s10 t2 + 9680s9 t3 13365s8 t4 + 15840s7 t5
20724s6 t6 + 9504s5 t7 + 8415s4 t8 16720s3 t9 + 6930s2 t10 1776st11 + 701t12
9.7
Mordells conjecture
After seeing a good many particular examples one might wonder whether anything is known about diophantine equations in general. For a long time only one
result in such a direction was known.
Theorem 9.7.1 (C.L.Siegel 1929) Let P (X, Y ) Z[X, Y ] be a polynomial,
irreducible in C[X, Y ]. Suppose that the genus of the projective curve given by
P = 0 is at least 1. Then P (x, y) = 0 has at most nitely many solutions in
x, y Z.
The proof is quite dicult and involves ideas from diophantine approximation
and arithmetic algebraic geometry. Standard examples of curves of genus 2 are
the hyper-elliptic curve y 2 = q(x), where q(x) is a polynomial of degree at least
5 and distinct zeros and the Fermat curve xn + y n = 1 with n > 3.
Already in 1922 L.J.Mordell conjectured that under the conditions of Siegels
theorem P (x, y) = 0 has at most nitely many solutions in x, y Q. This
conjecture withstood attempts to solve it for a long time until in 1983 G.Faltings
managed to provide a proof of it. Unfortunately this proof can only be understood
by experts in arithmetic algebraic geometry. In 1988 P.Vojta found a brilliant
new proof which, unfortunately, had the same drawback as Faltings proof in
that it was accessable only to a very small group of experts. In 1990 E.Bombieri
considerably simplied Vojtas proof, thus making it understandable for a large
audience of number theorists and algebraic geometers.
About polynomial diophantine equations in more than two variables almost nothing is known, although there exist a good many fascinating conjectures about
them.
9.8
Exercises
Exercise 9.8.1 Let a, b, c be any integral solution of a2 + b2 = c2 . Prove that 5

divides abc.
106
Exercise 9.8.2 Solve the equation x2 + y 2 = z 4 in x, y, z N with gcd(x, y, z) =

1.
Exercise 9.8.3 Let a, b, c be any solution of a2 + b2 = c4 with gcd(a, b, c) = 1.
Prove that 7 divides abc.
Exercise 9.8.4 Solve the equation x2 + y 2 = z 3 in x, y, z N with gcd(x, y, z) =
1. (Hint: use factorisation in Z[i]).
Exercise 9.8.5 (*) Prove that x4 y 4 = z 2 has no solution x, y, z N.
Exercise 9.8.6 Show that 2k 3l = 1, k, l 2 has only the solution k =
3, l = 2. (The conjecture that xk y l = 1, k, l 2 has only xk = 32 , y l = 23 as
solution has long been known as Catalans conjecture. In 2002 it was proven by
Michailescu).
Exercise 9.8.7 Prove that there exist no rectangular triangles with integral edges
whose surface area is a square.
Exercise 9.8.8 Solve 4y 2 = x3 + 1 in x, y Z.
Exercise 9.8.9 Prove that the only integer solution of
x2 + y 2 + z 2 = 2xyz
is x = y = z = 0.
Exercise 9.8.10 Prove that 32 1 is divisible by 2n+2 . Construct triples ak , bk , ck
N for k = 1, 2, . . . such that ak +bk = ck , gcd(ak , bk , ck ) = 1 and limk ck /N (ak bk ck ) =
. Here N (x) is the product of the distinct prime divisors of x.
n
Exercise 9.8.11 Suppose the abc-conjecture holds. Prove that there exist at most
nitely many triples ar , bs , ct such that ar + bs = ct , gcd(a, b) = 1 and 1/r + 1/s +
1/t < 1. (Hint: prove and make use of the following statement: 1/r + 1/s + 1/t <
1 1/r + 1/s + 1/t 1 1/42.)
Exercise 9.8.12 Find a, b, c Z such that 7 |abc and a7 + b7 c7 (mod 73 ).
Exercise 9.8.13 Show, assuming the abc-conjecture, the modied Hall conjecture which reads as follows. For every a < 1/2 there exists c(a) > 0 such that for
any positive integers x, y with x3 = y 2 we have |x3 y 2 | > c(a)xa .
Chapter 10
Prime numbers
10.1
Introductory remarks
In previous chapters we have seen that prime numbers play a crucial role in
number theory. We repeat here that by a prime number we mean an integer
which cannot be written as a product of smaller numbers. In this chapter we
occupy ourselves with the distribution of primes in N.
Theorem 10.1.1 (Euclid) There exist innitely many primes.
Proof. We have already seen Euclids proof in Theorem 1.4.5. Here we present
another proof due to Euler. Although slightly more complicated than Euclids
proof it uses an idea which will return repeatedly.
We shall show that
)
(
1
1
p
pN
tends to zero as N . Here, the product is taken over all primes p N .
Notice that by the unique factorisation theorem in Z,
)1 (
)
N
(
1
1
1
1
1
=
1 + + 2 + >
p
p p
n
n=1
pN
pN
Since the latter sum tends to innity as N we see that our product tends
to zero as N . Hence there are innitely many primes.
2
As a consequence of Eulers method we nd the following corollary.
1
Corollary 10.1.2 The sum
, taken over all primes, diverges.
p
1
Proof. Notice that for any x (0, 12 ) we have x > 12 log 1x
. Hence
1
1
1
> log
p
2
1 p1
pN
pN
107
108
CHAPTER 10. PRIME NUMBERS
and since the product tends to as N we see that our sum diverges.
A more precise analysis, as performed by Mertens, reveals that there exists a real
number A such that for all X > 2 we have
1
= log log X + A + O((log X)1 )
p
p<X
where the summation is over all primes p < X
Definition 10.1.3
(x) = #{p x| p prime}.
The local distribution of prime numbers seems to be completely erratic, and not
much is known about it. Remarkably enough one can say quite a few things about
the global distribution of primes as reected by (x). In the following table we
have counted s, the number of primes in the interval [x 75000, x + 75000] for
several values of x,
x
s
150000/ log x
8
10
8154
8143
109
7242
7238
10
10
6511
6514
11
10
5974
5922
1012
5433
5428
13
10
5065
5011
1014
4643
4653
15
10
4251
4342
The last column of this table suggests that the density of the primes near x is
about equal to 1/ log x. This led Gauss to conjecture,
x
dt
(x) li(x) :=
.
2 log t
The sign must be interpreted as asymptotic equality. More precisely, f (x)
g(x) means that f (x)/g(x) 1 as x . Since li(x) x/ log x we might also
conjecture,
x
(x)
.
log x
The rst result in this direction was obtained by Chebyshev who proved around
1852 that
x
x
< (x) < 1.11
0.92
log x
log x
for suciently large x. In 1860 Riemann, in a now historical paper, introduced
the complex function
1
(s) =
Res > 1
ns
n1
10.1. INTRODUCTORY REMARKS
109
in the study of prime numbers. For Res > 1 it is easy to see that (s) is analytic
and Riemann showed that it can be continued analytically to C with the exception
of s = 1, where it has a rst order pole with residue 1. The relation of (s) with
prime numbers is made apparent by the Euler factorisation
(s) =
(
p
1
1 s
p
)1
where the product is over all primes p. It turns out that for the distribution
of the primes the zeros of (s) in the critical strip 0 Res 1 are extremely
important. If there would have been no zeros in this strip one could have proved
some marvelous theorems on prime numbers. Unfortunately there are zeros in the
strip but, and this is fortunate again, they all seem to lie on the line Res = 1/2.
This was posed by Riemann as a question and until now no one has been able to
conrm it. This question, known as the Riemann hypothesis , is one of the classical
problems in mathematics. It follows very easily from Riemanns work that the
zeros lie symmetrical around the X-axis. So it suces to look at zeros in the
upper half plane. We order them in the order of increasing imaginary part. Then
it is known that the rst 1500 000 000 zeros are all simple and lie on Res = 1/2
(Brent, v.d.Lune, te Riele, Winter). It was proved by Levinson (1974) that at
least a third of the zeros in the critical strip is on the line Res = 1/2. It should
be noted that the zeros outside the critical strip are given by s = 2, 4, 6, . . .
and they are called the trivial zeros.
Continuing Riemanns work Hadamard and De la Vallee-Poussin, independently
of each other, proved the following theorem in 1896.
Theorem 10.1.4 (Prime number theorem)
(x)
x
.
log x
In 1949 Selberg and Erdos, more or less independently, gave an elementary proof
of the prime number theorem. By elementary we mean that no use is made of
complex function theory. It does not imply that the proof is simple!
Later it turned out that Gauss function li(x) is a better approximation of (x)
than x/ log x. Assuming the Riemann hypothesis we have
(x) = li(x) + O(x1/2 log x).
However, one is a long way o at proving such results. The best known estimate
are of the form
1
(x) = li(x) + O(x exp(
log x).
15
110
Here we give a table which compares values of (x) and li(x) for several x.
x
(x)
li(x) (x)
102
25
5
3
10
168
10
4
10
1229
17
105
9592
38
6
10
78498
130
107
664579
339
8
10
5761455
754
9
10
50847534
1701
1010
455052511
3104
11
10
4118054813
11588
1012
37607912018
38263
13
10
346065536839
108971
14
10
3204941750802
314890
1015 29844570422669
1052619
1016 279238341033925
3214632
As a peculiarity note that the values of (x) = li(x) (x) in our table are
all positive. Since x becomes quite large in our table one might suspect that
(x) is always positive. That one should be careful in stating such beliefs was
shown by Littlewood who proved in 1914 that (x) changes sign innitely often.
Around 1933 Skewes showed that, assuming the Riemann hypothesis, the rst
1034
sign change should be somewhere below 1010 . Numbers of this size were soon
called Skewes numbers. We now know, without having to assume the Riemann
hypothesis, that the rst change of (x) is somewhere below 6.7 10370 .
Finally a few words about the local distribution of primes. Let us denote by
p1 , p 2 , . . . , pn , . . .
the sequence of prime numbers in increasing order. First of all, pn+1 pn may
become arbitrarily large, i.e. there exist arbitrarily long gaps in the sequence
of prime numbers. To nd a gap of length at least N 1, say, we just have
to write down N ! + 2, N ! + 3, . . . , N ! + N and notice that these numbers are
divisible by 2, 3, . . . , N . By means of elementary methods Chebyshev proved
in 1852 that pn+1 < 2pn , thereby conrming Bertrands postulate . It is now
1
384
known, using deep analytic methods that pn+1 pn = O(pn ) with = 11
12
(Iwaniec, Pintz, Mozzochi). If Riemanns hypothesis is true then one can prove
that pn+1 pn = O( pn ). However, based on heuristic arguments, one expects

more, namely pn+1 pn = O((log pn )2 ) (Cramer).
If pn+1 = pn +2 then we call the pair pn , pn+1 a twin prime . Examples are (11, 13)
(29, 31) (41, 43) (71, 73) . . .. It is not known whether there exist innitely many
10.2. ELEMENTARY METHODS
111
twin primes, although one generally expects the answer to be yes.

The rst nontrivial result was by V.Brun in 1919 who showed that the series p1 , taken over
all primes belonging to a twin prime, converges. This was the rst example of
an important tool in analytic number theory, namely sieve methods. The largest
known twin prime in 2001 (September) was
318032361 2107001 1
but such records usually have a short life.
For much more information we refer to P.Ribenboims delightful The book of
prime number records.
10.2
Elementary methods
In this section we shall prove two theorems by elementary methods which are
similar in spirit to the methods used by Chebyshev.
Theorem 10.2.1 Let n N and n > 10. Then
1 n
n
< (n) < 3
.
3 log n
log n
Theorem 10.2.2 (Bertrands postulate) For any n N there exists a prime
number p such that n .
n
n
)
(
)
2n
2n(2n 1) 2(n 1)
=
.
n
nn
n1
Since 4(n 1)/n < 2n(2n 1)/(n n) < 4 this implies

(
) ( )
(
)
n 1 2(n 1)
2n
2(n 1)
4
<
<4
.
n
n1
n
n1
(10.1)
To prove statement (a) notice that it is true for n = 5 and apply induction on n
using the second inequality in (10.1). To prove (b) note that it is true for n = 4
and apply induction on n using the rst inequality in (10.1).
2
112
Lemma
( ) 10.2.4 Let p be prime and n N. Let k be the number of prime factors
p in 2n
. Then
n
(a) pk 2n.
(b) If n 2 then k = 0.
Proof. For any m N we know that the number of factors p in m! is equal to
[
] [ ]
]
[
m
m
m
+ 2 + =
.
p
p
pr
r=1
Hence
k=
]
([
2n
pr
r=1
n
2 r
p
])
.
(10.2)
Notice that 0 [2x] 2[x] 1 for any x R. Notice also that the terms in
(10.2) vanish when r > log 2n/ log p. Hence
k
rlog 2n/ log p
]
log 2n
log 2n
1=
log p
log p
and assertion (a) follows immediately.

If n 2n and all terms in (10.2) with r > 1
vanish. Hence k = [2n/p] 2[n/p]. But 2n/3 < p n implies 1 n/p < 3/2
and thus we nd that k = 0, as asserted in (c).
2
Lemma 10.2.5 For any n 2 we have
p < 4n .
pn
pprime
( )
Proof. Let m N and m 5. Observe that according to Lemma 10.2.4(b) 2m
m
is divisible by all primes p with m < p 2m. Hence, using Lemma 10.2.3(a),
p < 4m1 .
m<p2m
(10.3)
10.2. ELEMENTARY METHODS
113
We shall prove our statement by induction on n. First of all our lemma can be
veried byhand forall n < 10. Now suppose n 10. If n is even we have
obviously pn p = pn1 p < 4n1 < 4n and we are done. If n is odd we write
p=
p
p.
p n+1
2
pn
n+1
<pn+1
2
The rst product on the right can be estimated using our induction hypothesis
and the second by using (10.3) with m = (n + 1)/2. We obtain pn p <

4(n+1)/2 4(n+1)/21 = 4n , as desired.
2
Proof of Thm 10.2.1. First we prove the upper bound. Choose R between
0 and 1. Notice that Lemma 10.2.5 implies that
p < 4n .
n<pn
Since each factor in the product is larger than
(1/2)((n)( n))
n this implies that
< 4n
and hence
n
(n) ( n) < 2 log 4
.
log n
If we estimate ( n) by n/2 we nd that

(n) < 2.8
n
n0.5
+
.
log n
2
When n > 200 this can be bounded by 3n/ log n as desired. For n 200 we have
to verify this upper bound case by case.
Let m 4. By combination of Lemma 10.2.3(b) and Lemma 10.2.4(a) the lower
bound is derived as follows
( )
4m
2m
<
(2m) = (2m)(2m) .
m
m
p2m
Hence
log(4m /m)
2m
> log 2
1.
log 2m
log 2m
Suppose n 8. When n = 2m is even we deduce,
n
(n) log 2
1
log n
(2m)
and when n = 2m + 1 is odd we have

(n) = (2m + 1) = (2m + 2) > log 2
2m + 2
n
1 > log 2
1.
log(2m + 2)
log n
When n 10 the bound log 2 logn n 1 can be bounded below by
1 n
3 log n
as desired. 2
114
Proof of Thm 10.2.2. Using Lemma 10.2.4(a)(b)(c) we see that

( )
2n
p
p
<
(2n).
n
n<p2n
p2n/3
(10.4)
p 2n
The second product on the right can be estimated using Lemma 10.2.5 and yields
the upper
bound 42n/3 . The third product on the right can be estimated by
2n
(2n) . Elementary estimates show that this can be bounded above by 4n/3 /n
as soon as n > 512. Hence when n > 512 we obtain from (10.4),
( ) ( ) n
2n
4
 4n /n. Together with
From Lemma 10.2.3(b) we have thelower bound 2n
n
(10.5) this implies that the product n<p2n p is non-empty when n > 512. For
n 512 observe that
2, 3, 5, 7, 13, 23, 43, 83, 113, 223, 443, 881
is a sequence of prime numbers of which each term is smaller than twice its
predecessor. Hence the theorem is also true for n 512.
2
10.3
Exercises
Exercise 10.3.1 Prove,

a)
(1 ps )1
ns
b)
s 1
(1 p )
<
ns
(1 ps )1 =
ns
s > 1.
n=1
p prime
Exercise 10.3.2 Prove,

a)
(1 ps )1
ns
pX
p prime
s > 1, X 1
n=1
pX
p prime
c)
s R>0 , X 1
nX
pX
p prime
nX
s R>0 , X 1
10.3. EXERCISES
115
b)
1
> log X
n
nX
X 1.
Now choose s = 1 in part a), take logs on both sides and show that
c)
1
1
> log log X .
p
2
pX
p prime
Exercise 10.3.3 a) Prove that

(
)
x
dt
x
=O
k
(log x)k
2 (log t)
k N, x 2.
(Hint: split the integration interval into two parts.)

b) Prove that
x
x
x
+ 1!
+ + (k 1)!
+O
li(x) =
2
log x
(log x)
(log x)k
x
(log x)k+1
)
n N.

( p2 + 1 ) 5
= .
p2 1
2
p prime
(Hint: use (2) = 2 /6 and (4) = 4 /90).
Exercise 10.3.5 Let p1 , p2 , p3 , . . . be the sequence of consecutive prime numbers.
Prove, using the prime number theorem, that pn /n log n 1 as n .
Exercise 10.3.6 Let > 0. Prove, using the prime number theorem, that there
exists x0 () such that for any x > x0 the interval [x, (1 + )x] contains a prime
number. (The case = 1 is known as Bertrands postulate.)
Exercise 10.3.7 Verify, using the prime number theorem, whether or not the
following sums converge,
1
,
p
log
p
prime
log p
.
p
p prime
Exercise 10.3.8 Prove, using the prime number theorem, that

log(lcm(1, . . . , n))
= 1.
n
n
lim
116
Exercise 10.3.9 (*) (H.W.Lenstra jr.) Prove that for innitely many n N we
have (n)|n. (Examples: (30) = 10|30, (1008) = 168|1008).
1
Exercise 10.3.10 Consider for any n the integral In = t=0 tn (1 t)n dt.
a)Prove that In is a rational number whose denominator divides lcm[n, . . . , 2n+1].
(Hint: integrate term by term).
b)Prove that |In | < (1/4)n .
c)Prove that lcm[1, . . . , m] 2m1 for all m N.
Exercise 10.3.11 Use the ideas of the
previous exercise.
a) Prove that |t(1 t)(1 2t)| < 1/6 3, t [0, 1].
b) Prove that lcm(1, 2, . . . , 6n + 1) 108n n N.
1
c) Prove that lcm(1, 2, . . . , m) 108
((108)1/6 )m , m N.
Chapter 11
Irrationality and transcendence
11.1
Irrationality
Definition 11.1.1 Let C. We call irrational when Q.

Proving irrationality and transcendence of numbers is now being considered as a
branch of number theory, although the techniques that are used involve subjects
like complex analysis, linear dierential equations and algebraic geometry. In this
chapter we shall restrict ourselves to examples in which only elementary methods
are used.
The easiest numbers for which irrationality can be proved are the algebraic numbers.
Theorem 11.1.2 Let be a zero of a polynomial xm + c1 xm1 + + cm Z[x].
Then is either irrational or Z. In the latter case we have |cm .
Proof. Suppose xm + c1 xm1 + + cm has a rational zero p/q with p, q
Z, (p, q) = 1, q > 0. Then pm + c1 pm1 q + + cm q m = 0. Hence q|pm and since
(p, q) = 1 this implies q = 1. Notice that pm + c1 pm1 + + cm = 0 now implies
that p|cm .
2
Corollary11.1.3 Let m N. If N Z is not the m-th power of an integer
then = m N is irrational.
Proof. Notice that m N = 0. If Q then Z according to Theorem 11.1.2 and hence N is the m-th power of an integer. This contradicts our
assumptions, hence Q.
2
Theorem 11.1.4 Let e be the base of the natural logarithm. Then e is irrational.
117
118
CHAPTER 11. IRRATIONALITY AND TRANSCENDENCE
Proof. Suppose e is rational with denominator d. We use the series expansion

e=
1
.
k!
k=0
The number
1
1
1

1! 2!
k!
is a positive rational number with a denominator dividing d(k!). Hence, since
is not zero,
1
.
k!d
On the other hand,
:= e 1
1
1
+
+
(k + 1)! (k + 2)!
(
)
1
1
1
+
+
=
k! k + 1 (k + 1)(k + 2)
(
)
1
1
1
<
+
+
k! k + 1 (k + 1)2
11
=
k! k
contradicting our lower bound for whenever k > d.
The irrationality proof of is more complicated and we require the following

lemma.
Lemma 11.1.5 Let m Z0 . Then
1
tm sin tdt
0
is a polynomial in 1/ 2 with integral coecients and degree [m/2].

Proof. By induction on m. For m = 0 and m = 1 we have
1
1
sin tdt = 2
t sin tdt = 1.
0
Suppose m > 1. After a two-fold partial integration we obtain

1
1
m(m 1)
m
tm2 sin tdt
t sin tdt = 1
2
0
0
from which our assertion follows.
11.1. IRRATIONALITY
119
Theorem 11.1.6 We have 2 Q and Q.

Proof. Dene Pm (t) =
the integral
1 d m m
( ) t (1t)m
m! dt
and notice that Pm (t) Z[t]. Consider
In =
(sin t)P2n (t)dt.

0
After a 2n-fold partial integration we nd

2n
In = (1)
(2n)!
(sin t)t2n (1 t)2n dt.
Hence
2n+1
0 < |In | <
.
(2n)!
On the other hand we can compute In term by term. Lemma 11.1.5 then implies
that In = An (1/ 2 ), where An (x) Z[x] and deg An n.
Suppose that 2 = a/b for some a, b N. Because An has degree n the number
In = An (1/ 2 ) = An (b/a) is a fraction whose denominator divides an . Moreover,
In = 0. Hence
1
|In | n .
a
Together with the upper bound for |In | this implies
1
2n+1
<
a2n
(2n)!
which becomes impossible when n . Hence 2 Q. This immediately
implies Q.
2
Around 1740 Euler proved e to be irrational and the rst proof of the irrationality
of was given by Lambert in 1761. This proof was based on the continued fraction
expansion of arctg(x). The proof we gave can be considered as a variation of a
proof given by I.Niven. On the other hand there are many naturally occurring
numbers for which no irrationality results are
known. For example, it is not
known whether Eulers constant = limn ( nk=1 (1/k) log n) or e + or e
is irrational. Motivated by the standard series for e P.Erdos asked the following
question. Is
1
k! + 1
k=1
irrational? Surprisingly this seems to be dicult to answer.
120
11.2
Transcendence
Definition 11.2.1 A number C is called algebraic if it is the zero of a nontrivial polynomial with coecients in Q. A number is called transcendental if it
is not algebraic.
Obviously, proving transcendence of a number is much harder than proving irrationality. It is therefore no surprise that in the beginning of the 19th century no
examples of transcendental numbers were known. In 1844 Liouville proved the
following theorem.
Theorem 11.2.2 (Liouville 1844) Let be an algebraic number whose minimal polynomial has degree n. Then there exists a positive constant c such that
for any p, q Z and q > 0,

p
c .

q qn
Proof. Let P (x) Z[x] be the minimal polynomial of . Since P has no rational
roots we have P (p/q) = 0 and, moreover, |P (p/q)| 1/q n . Write P (x) =
(x )Q(x) and let M = max|x|1 |Q(x)|. Suppose | p/q| 1. Then,
trivially, |P (p/q)| M |p/q|. Combined with our lower bound for |P (p/q)| this
yields | p/q| 1/(M q n ). This proves our theorem with c = min(1, 1/M ). 2
Corollary 11.2.3 Let R. Suppose that there exists a sequence of rational
numbers {pn /qn }
n=1 , with qn > 0 for all n, and a sequence of numbers n such
that

pn
c

0 < < n
lim n =
n
qn
q
n
for some c > 0. Then is transcendental.

This corollary enabled Liouville to construct innitely many examples of transcendental numbers. We give one example here,leaving the construction of other
examples to the reader.
Corollary 11.2.4 The number
1
=
2k!
k=0
is transcendental.
Proof. We apply Corollary 11.2.3. Let qn = 2n! and pn = 2n! nk=0 (1/2k! ). Then

1
p
1 1
2
2
n
=
= (n+1)! = n+1 .
< (n+1)!

k!
j

qn
2 2
2
2
qn
j=0
k=n+1
11.2. TRANSCENDENCE
121
Application of Corollary 11.2.3 yields our result.
Through the pioneering work of Cantor on set theory around 1874 it also became
clear that almost all real numbers are transcendental. This follows from the
following two theorems.
Theorem 11.2.5 The set of algebraic numbers is countable.
Proof. It suces to show that the set Z[X] is countable. To any polynomial
P (X) = pn X n + pn1 X n1 + + p1 X + p0 Z[X] with pn = 0 we assign the
number (P ) = n + |pn | + |pn1 | + + |p0 | N. Clearly for any N N the
number of solutions to (P ) = N is nite, because both the degree and the size
of the coecients are bounded by N . Hence Z[X] is countable.
2
Theorem 11.2.6 (Cantor) The set of real numbers is uncountable.
Proof. We will show that the set of real numbers in the interval [0, 1) is uncountable. Suppose that this set is countable. Choose an enumeration and
denote the decimal expansion of the n-th real number by 0.an1 an2 an3 , where
anm {0, 1, 2, 3, 4, 5, 6, 7, 8, 9} for all n, m. Now consider the real number whose
decimal expansion reads 0.b1 b2 b3 where the bi are chosen such that bi = aii for
every i. This choice implies that does not occur in our enumeration. Hence
[0, 1) is uncountable.
2
The principle of the proof of Theorem 11.2.6 is known as Cantors diagonal procedure and it occurs in many places in mathematics.
Almost all real numbers being transcendental, it seems ironic that until the end
of the 19-th century not a single naturally occurring number was known to be
transcendental. Only in 1873 Hermite showed that e is transcendental and in
1882 Lindemann proved to be transcendental. In his famous lecture of 1900
D.Hilbert asked whether numbers of the form ab with a, b algebraic, a = 0, 1
and b Q, are transcendental. Specic examples are 2 2 and i2i = e . This
problem was considered to be very dicult by Hilbert, but already in the 1930s
A.O.Gelfond and Th.Schneider indepently developed techniques to solve this
problem. So now we know,
Theorem 11.2.7 (Gelfond, Schneider ,1934) Let a, b be algebraic and suppose that a = 0, 1 and b Q. Then ab is transcendental.
Corollary 11.2.8 Let , be two positive real algebraic numbers such that = 1
and log / log Q. Then log / log is transcendental.
Proof. Let b = log / log and suppose b is algebraic. Then, according to
Theorem 9.2.7 the number = b is transcental which is impossible since is
algebraic.
2
Nowadays the Gelfond-Schneider theory has grown into a eld of its own in which
large classes of numbers, ususally related to algebraic geometry, are known to be
transcendental.
122
11.3
Irrationality of (3)
s
Let (s) =
n=1 n . We shall be interested in the numbers (m) with m Z2 .
Euler showed that
B2n 2n
(2n) = (1)n1 22n1
(2n)!
where B2n is the 2n-th Bernoulli number. Hence (2n) is transcendental because
2n is transcendental. Strangely enough next to nothing is known about the
numbers (2n + 1) for n 1. It was therefore a complete surprise when in
1978 the french mathematician R.Apery announced a proof of (3) Q. The
rst reaction of his fellow mathematicians was incredulity, since the presentation
of the proof was a mixture of remarkable formulae and downright impossible
statements. Later this proof was patched up by H.Cohen and D.Zagier and
Apery turned out to be correct on all the crucial parts. The simple proof we
present here was found by F.Beukers, but the shape of the integrals is motivated
by Aperys formulae.
Lemma 11.3.1 Let r, s Z0 . If r > s then
Z
log xy r s
x y dxdy
.
1 xy
[1, 2, . . . , r]3
If r = s then
(
)
log xy r r
1
1
x y dxdy = 2 (3) 3 3 .
1 xy
1
r
Proof. Let > 0 and consider the integral
xr+ y s+
dxdy.
1 xy
Develop (1 xy)1 in a geometrical series and carry out the integration term by
term. We nd
1
.
(k + r + + 1)(k + s + + 1)
k=0
When r > s this implies
(
)
xr+ y s+
1
1
1
dxdy =
1 xy
rs k+s++1 k+r++1
k=0
(
)
1
1
1
+ +
=
rs s+1+
r+
11.3. IRRATIONALITY OF (3)
123
Dierentiate with respect to and put = 0,

(
)
1 1
log xy r s
1
1
1
Z
x y dxdy =
+ + 2
.
2
r s (s + 1)
r
[1, 2, . . . , r]3
0
0 1 xy
When r = s we nd in a similar way,
)
(
log xy r r
1
2
1
x y dxdy =
= 2 (3) 3 3
3
1 xy
(k
+
r
+
1)
1
r
k=0
2
Lemma 11.3.2 When n is suciently large, lcm[1, 2, . . . , n] < 3n .

Proof. Notice that
lcm[1, 2, . . . , n] =
p[log n/ log p] <
pn
plog n/ log p
pn
n = n(n)
pn
where the products are taken over the primes p. According to the prime number
theorem we have (n) < (log 3)n/ log n for suciently large n. Hence n(n) < 3n
for n suciently large.
2
Theorem 11.3.3 (R.Ap
ery 1978) The number (3) is irrational.
Proof. Consider the double integral
1 1
log xy
In =
Pn (x)Pn (y)dxdy
1 xy
0
0
( d )n n
where Pn (x) = n!1 dx
x (1 x)n . Notice that P (x) Z[x]. From Lemma 11.3.1
it follows that
An + Bn (3)
In =
,
An , Bn Z.
[1, 2, . . . , n]3
Notice that
log xy
=
1 xy
hence
In =
1
dz
1 (1 xy)z
Pn (x)Pn (y)
dxdydz
1 (1 xy)z
111
where stands for 0 0 0 . After an n-fold partial integration with respect to
x we obtain
(xyz)n (1 x)n Pn (y)

In =
dxdydz.
(1 (1 xy)z)n+1
124
Substitute z = (1 (1 xy)w)/(1 w). Then
Pn (y)
dxdydw.
In = (1 x)n (1 w)n
1 (1 xy)w
An n-fold partial integration with respect to y yields
n
x (1 x)n y n (1 y)n wn (1 w)n
In =
dxdydw.
(1 (1 xy)w)n+1
It is an excercise to show that for all 0 x, y, w 1 we have

x(1 x)y(1 y)w(1 w)

( 2 1)4 .

1 (1 xy)w
Hence
|In | < ( 2 1)
4n
dxdydw
= 2( 2 1)4n (3).
1 (1 xy)w
On the other hand, In = 0 and In = (An + Bn (3))/[1, 2, . . . , n]3 . Suppose

(3) = p/q with p, q Z and q > 0. Then, using Lemma 11.3.2,
1
1
<
|In | < 2( 2 1)4n (3).
n
3
27 q
[1, 2, . . . , n] q
Hence,
1 < ( 2 1)4n 27n 2(3)q.
Since ( 2 1)4n 27n tends to 0 as n , we have a contradiction. Therefore,

(3) Q.
2
11.4
Exercises
Exercise 11.4.1 Prove, using the series expansions for e and e1 , that e is not
algebraic of degree 2.
Exercise 11.4.2 Show that
log 3
log 2
is irrational.
Exercise 11.4.3 Prove that e + and e are not both rational.

( )n
4
1
Q.
5
3n2
n=0
Exercise 11.4.5 Show that the so-called Champernowne number

0.1234567891011121314151617 . . .
is irrational (Actually K.Mahler proved it transcendental in the 1930s, but that
is much harder).
Chapter 12
Solutions to selected problems
(1.5.3) When m is odd we have the polynomial factorisation
xm + 1 = (x + 1)(xm1 xm2 + x + 1).
Suppose that n is not a power of 2, i.e. n contains an odd divisor m > 1.
Suppose n = m k. Substitute x = 2k in the above identity, then
2n + 1 = 2mk + 1 = (2k + 1)(2k(m1) 2k(m2) + 2k + 1).
So 2k +1 is a divisor of 2n +1. It remains to point out that it is a non-trivial
divisor, i.e. 1 < 2k + 1 < 2n + 1. So 2n + 1 is not prime, a contradiction.
Therefore n cannot contain odd divisors > 1.
(1.5.2) When m is odd we have the polynomial factorisation
xm 1 = (x 1)(xm1 + xm2 + + x + 1).
Suppose that n is not a prime i.e. n = m k for some integers k, m > 1.
Substitute x = 2k in the above identity, then
2n 1 = 2mk 1 = (2k 1)(2k(m1) + 2k(m2) + + 2k + 1).
So 2k 1 is a divisor of 2n 1. It remains to point out that it is a non-trivial
divisor, i.e. 1 < 2k 1 < 2n 1. So 2n 1 is not prime, a contradiction.
Therefore n cannot be composite.
(1.5.6) Suppose that p + 2 is composite for only nitely many primes p. Then there
is a number P0 such that p prime and p > P0 implies p + 2 prime. Choose
a prime p > P0 . Then, consequently, all odd numbers larger than p are
prime. This is clearly impossible since, fore example, all powers of 3 are
composite. We get a contradiction.
125
126
CHAPTER 12. SOLUTIONS TO SELECTED PROBLEMS
(1.5.8) Since n is odd we can write n = 2m + 1. Hence

(
)
m+1
n = (2m + 1) = 4m + 4m + 1 = 8
+ 1.
2
2
So we see that n2 is 1 modulo 8.

(1.5.12) For a),b) we refer to the course notes. Let eta = (1 +
ratio. Note that 2 = + 1.
5)/2, the golden
Part c) When n = k we must prove that rk 1 which is clearly true, since

rk is the last non-zero remainder. When n = k 1 we must show that
rk1 which is true as well, since rk1 2. Now we apply induction using
rn2 =
qn1 rn1 + rn
rn1 + rn
kn+1 + kn = kn ( + 1)
kn 2 = kn+2
Part d) We have a = r1 and r1 k+1 . Hence k+1 a, from which our

desired inequality follows.
(1.5.14) We give a hint. Suppose n > m and suppose n = mq + r with 0 r < m.
Notice that we have the polynomial identity
(xn 1) = (xm 1)(xnm + xn2m + + xnqm ) + xr 1.
By application of the euclidean algorithm to an 1 and am 1 we see
that the exponents are precisely the remainders of the euclidean algorithm
applied to n, m.
(1.5.16) Let k be the number of zeros. In other words, k is the highest power such
that 10k divides 123!. Since there are more factors 2 than 5 in 123!, it
suces to count the number of factors 5. Between 1 and 123 there are
24 ve-tuples and 4 twentyve-tuples (twentyve-tuples are also counted
as ve-tuples). There are no hundredtwentyve-tuples or higher between 1
and 123. In total this gives us 24+4 factors 5. Hence k = 28.
(1.5.17(a)) Consider the numbers 1 to n. Among these numbers there are [n/p] ptuples, [n/p2 ] p2 -tuples, [n/p3 ] p3 -tuples, etc. Here we count pk -tuples also
as pk1 -tuples. The total number of factors p in n! therefore equals
[ ] [ ] [ ]
n
n
n
+ 2 + 3 +
p
p
p
127
(2.3.1) A perfect number n is characterised by (n) = 2n. In other words,
2n divide on both sides by n to obtain
d|n
d=
d
= 2.
n
d|n
Notice that as d runs over the divisors of n, the number d/n runs over all
inverses of the divisors of n. Hence
1
= 2.
d
d|n
(2.3.3) All convolution products are convolution products of multiplicative functions. Hence the convolution products are also multiplicative. To describe
the products it suces to describe their values in the prime powers. Here
are the results,
1. (Ik Ik )(n) = 0 (n)nk
2. I1 =
3. ( )(pk ) = 2 als k = 1, 1 als k = 2, 0 als k > 2
4. ( 2 )(n) = |(n)|
5. (2 2 )(pk ) = 4k
6. ( )(pk ) = pk 2pk1 + pk2 if k 2, p 2 if k = 1.
(2.3.7) We shall prove something more general. Let f (n) be any arithmetic function
with f (1) = 0. Then there exists an arithmetic function g such that f g =
e. In other words, we must determine numbers g(1), g(2), g(3), . . . such that
f (1)g(1) = 1 and for all n > 1,
f (d)g(n/d) = 0
d|n
Hence g(1) = 1/f (1) and for n = 2, 3, 4, . . .,
f (1)g(n) =
g(d)f (n/d).
d|n,d<n
Note that the latter relation allows us to determine g(2), g(3), g(4), . . . recursively. Notice also that g is uniquely determined.
To complete our exercise, we must show that if f is multiplicative, then
so is g. To that end we construct a multiplicative function h such that
f h(pk ) = 0 for all prime powers pk and f h(1) = 1. By the multiplicative
128

property of f, h it follows that f h = e and by the uniqueness of g we
conclude that g = h. For the construction of h we use the analogue of the
construction for g, but now restricted to prime powers pk ,
f (1)h(p ) = h(p ) =
k
k1
h(pl )f (pkl ).
l=0
The values of h at other integers are now dened by the requirement of

multiplicativity.
(2.3.8) Notice that both lefthand side and righthand side are multiplicative functions. Thus it suces to show equality when n is a prime power pk . Note
that
k
k
0 (d)3 =
0 (pl )3 =
(l + 1)3 .
d|pk
l=0
l=0
Also note that

k
k
l 2
2
0 (p )) = ( (l + 1))2 .
(
0 (d)) = (
d|pk
l=0
l=0
The two results are equal because we have the famous identity
13 + 23 + + (k + 1)3 = (1 + 2 + + (k + 1))2
for all positive integers k.
(3.5.1) Suppose that n = ab with 1 < a < b < n. Then the product (n 1)!
contains both factors a and b. Hence ab divides (n 1)!. Suppose n is
composite and suppose it cannot be written as a product of two distinct
numbers a, b < n. Then n must be the square of a prime p, i.e. n = p2 .
We have assumed n > 4, so p > 2. But in that case the product (n 1)!
contains the factors p and 2p. So p2 divides (n 1)!.
(3.5.3) Note that a number is divisible by 11 if and only if the alternating sum of its
digits is divisible by 11. The alternating sum of the digits of a palindromic
number of even length is zero.
(3.5.4) Answers: 3(mod 7), 13(mod 71), 83(mod 183).
(3.5.6) Answers:
a) x 1307(mod 2100)
b) y 675(mod 1540)
c) z 193(mod 420)
129
(3.5.8) Instead of a million consecutive numbers we can more generally ask for n
consecutive numbers, where n is any integer. We choose n distinct primes
p1 , p2 , . . . , pn and solve the simultaneous system of congruences
x + 1 0(mod p21 ) x + 2 0(mod p22 )
x + n 0(mod p2n ).
Since the numbers p2i are pairwise relatively prime, the Chinese remainder
theorem shows the existence of a solution x N. Hence x+1, x+2, . . . , x+n
is a sequence of n consecutive numbers all divisible by a square > 1.
(3.5.9) Notice that by the chinese remainder theorem,
a2 a(mod 10k ) a2 a(mod 2k ),
a2 a(mod 5k ).
Suppose we want to solve a2 a(mod pk ) for any prime p. Observe that

the equation is equivalent to pk |a2 a, hence pk |a(a 1). Since a and
a 1 are relatively prime, we have either pk |a or pk |a 1. In other words,
a 0(mod pk ) or a 1(mod pk ). Applying this to p = 2, 5 we get the
following possibilities
1. a 0(mod 2k ), a 0(mod 5k ). But this implies that a is divisible by
10k . This gives trivial solutions which are ruled out by the constraint
1 < a < 10k .
2. a 1(mod 2k ), a 1(mod 5k ). In this case a 1 is divisible by 10k ,
which is another trivial solution ruled out by the extra requirement
1 < a < 10k .
3. a 0(mod 2k ), a 1(mod 5k ). According to the chinese remainder
theorem this has a unique residue class modulo 10k as solution. It is
not 0 or 1 modulo 10k , hence there exists a unique solution a with
1 < a < 10k .
4. a 1(mod 2k ), a 0(mod 5k ). Again this gives a unique solution.
In all we nd two solutions to our general problem.
Now let k = 12 according to the above we must rst solve a 0(mod 212 ), a
1(mod 512 ), which gives us a 81787109376(mod 1012 ) and hence a =
81787109376. Secondly we must solve a 1(mod 212 ), a 0(mod 512 ),
which gives us a = 918212890625. These are the two solutions.
Notice that the sum of the non-trivial solutions found above, add up to
1000000000001. Can you explain that?
(3.5.12) Use the map a(mod 10) 7 (a(mod 2), a(mod 5) to nd
0(mod 10) 7 (0(mod 2), 0(mod 5))
130

1(mod
2(mod
3(mod
4(mod
5(mod
6(mod
7(mod
8(mod
9(mod
10)
10)
10)
10)
10)
10)
10)
10)
10)
7
7
(1(mod
(0(mod
(1(mod
(0(mod
(1(mod
(0(mod
(1(mod
(0(mod
(1(mod
2), 1(mod
2), 2(mod
2), 3(mod
2), 4(mod
2), 0(mod
2), 1(mod
2), 2(mod
2), 3(mod
2), 4(mod
5))
5))
5))
5))
5))
5))
5))
5))
5))
(3.5.13) Case a) We must show that 42

3(mod 13). Notice that the order of
4(mod 13) equals 6. This means that in the determination of 4k (mod 13)
for any k, only the value of k(mod 6) matters. So we must determine
22n+1 (mod 6) for all n. Note that 22n+1 0(mod 2) and 22n+1 (1)2n+1
2n+1
1(mod 3) for all n. Hence 22n+1 2(mod 6). We conclude that 42
42 3(mod 13).
2n+1
Case b) When 37 divides x, the statement is certainly true. Suppose now

that 37 does not divide x. In that case, x36 1(mod 37). So we must
determine 99 (mod 36). Notice that 99 0(mod 9) and 99 19 1(mod 4).
Hence 99 9(mod 36). A quick calculation shows that x9 (mod 37) has in
total 4 dierent values, 1, 1, 6, 6. After adding 4 to these numbers we
get 5, 3, 10, 2(mod 37) None of these numbers is 0(mod 37).
(3.5.14) By Eulers theorem we have
ki
)
a(pi
ki 1
(pi 1)
api
1(mod pki i ) for i = 1, 2, . . . , r
Hence
a(n) 1(mod pki i ) for i = 1, 2, . . . , r
By the Chinese remainder theorem this implies a(n) 1(mod n).
(3.5.15) Notice that 2730 = 2 3 5 7 13. By the Chinese remainder theorem
it suces to show that n13 n(mod p) for all n and p = 2, 3, 5, 7, 13. We
repeatedly use Fermat little theorem.
n13 n(mod 13) for all n (Fermats little theorem).
n13 n7 n6 n n6 n7 n
(mod 7).
n13 (n5 )2 n3 n2 n3 n5 n(mod 5).

n13 (n3 )4 n n4 n n3 n n n3 n(mod 3).
n13 n(mod 2) because n is even n13 is even.
131
(3.5.16) We must solve: 2p1 1 = pu2 in an odd prime p and an integer u. Note
that u must be odd. factorisation of the left hand side yields (2(p1)/2
1)(2(p1)/2 + 1) = pu2 . The factors on the left are relatively prime (check!).
This implies that there exist positive integers r, s such that either
2(p1)/2 1 = pr2 ,
2(p1)/2 + 1 = s2
or
2(p1)/2 1 = r2 ,
2(p1)/2 + 1 = ps2 .
In the rst case s2 1 is a power of 2. So s1 = 2k and s+1 = 2l for certain

integers k, l met l > k > 0. taking the dierence, 2l 2k = 2. From this
follows that 2k |2 and hence k = 1. Consequently, s = 3 and 2(p1)/2 +1 = 8.
This gives us p = 7. A small check, (26 1)/7 = 9, a square.
In the second case we see that r2 + 1 is a power of 2. Since r is odd, we
have r2 1(mod 4) and hence r2 + 1 2(mod 4). In other words, r2 + 1
contains at most one factor 2. So, r2 +1 = 2 r = 1 2(p1)/2 1 = 1.We
conclude that p = 3. A small check, (22 1)/3 = 1, a square.
(3.5.17) Wilsons theorem tells us that (p 1)! 1(mod p). We now rewrite the
product (p 1)! as
(
)
p1
p+1
(p 1)! =
!
(p 1).
2
2
The second group of factors on the right is modulo p equal to the factors
(p 1)/2, (p 3)/2, . . . , 2, 1. Hence
((
(p 1)! = (1)
(p1)/2
) )2
p1
! .
2
If we now notice that (1)(p1)/2 = 1 (because p 1(mod 4)) and (p1)!

1(mod p), our assertion follows.
(3.5.20) Notice that an 1(mod an 1). Furthermore, ak 1(mod an 1) for all
0 < k < n because 1 < ak 1 < an 1 for all such k. Hence the order of
a(mod an 1) is precisely n. The order of an element divides the order of
the group, hence n (an 1).
(3.5.21) This is slightly tedious. Write (n) = n p|n (1 1/p). We give a lower
bound for p|n (1 1/p). Notice that
p|n
(1 1/p)
1
2
(1 1/p)
p|n, p odd
132

The number of distinct primes dividing n can be at most log2 (n) (the logarithm of n in base 2). Note that for each odd prime p we have 11/p 2/3.
Hence
( )log2 (n)
1 2
(1 1/p)
2 3
p|n
1 log(2/3)/ log(2)
n
2
1 0.5
n
2
=
We conclude that (n)

n .
1
2
n10.5 =
1
2
n0.5 . The latter goes to as
(3.5.28) The orders are 6, 11, 8 respectively. To shortcut the computation, notice
for example that (46) = 22. So the order of 3(mod 46) divides 22. Thus
we need only check whether, 31 , 32 , 311 are 1(mod 46). If not, then 22 is the
order of 3(mod 46). It turns out that 311 1(mod 46).
(3.5.30) Note that 2p 1(mod q). The order of 2(mod q) thus divides p. Since
p is a prime and 21 1(mod q) the order is precisely p. Hence p divides
(q) = q 1. So q 1(mod p). Furthermore, since q is odd, q 1(mod 2).
Hence, because p is odd, we conclude that q 1(mod 2p). So q is of the
form q = 2mp + 1.
(3.5.31) To show part (a) notice that a2 1(mod q). Together with its square
n+1
a2
1(mod q) we note that a has order 2n+1 in (Z/qZ) . Hence 2n+1
divides q 1 which solves part (a).
n
We now follow a variation on Euclids proof. Suppose there are nitely

many primes p with p 1(mod 2n ). Call them p1 , p2 , . . . , pr . Let q be a
n
prime divisor of N = (2p1 p2 pr )2 + 1, which is necessarily odd. Then
it follows from (a) that q 1(mod 2n ). So there is an i such that q = pi .
Hence N 1 is divisible by q. Together with q|N this gives q|1 which is
impossible. We have a contradiction. There are innitely many primes of
the form k 2n + 1.
(3.5.33) Trial and error shows that 2 is a primitive root modulo 11. All other primitive roots modulo 11 can then be obtained by computation of 2k (mod 11) for
1 k < 10 and gcd(k, 10) = 1. So, 21 2(mod 11), 23 8(mod 11), 27
7(mod 11), 29 6(mod 11) are the primitive roots modulo 11.
From the theory it follows that 2(mod 112 ) has order 10 or 110. Checking
that 210 56(mod 121) we conclude that 2(mod 121) has order 110 and
hence is a primitive root. All other primitive roots modulo 121 can be
133
computed by computing 2k (mod 121) with 1 k < 110 and gcd(k, 110) =
1.
(3.5.34)
1. By testing 2d (mod 13) for all divisors d of 12 we quickly see that 2 is

a primitive root modulo 13. The other primitive roots are given by
2k (mod 13), where gcd(k, 12) = 1. So, 2, 6, 7, 11(mod 13).
Notice (Z/14Z) (Z/7Z) (Z/2Z) (Z/7Z) . The latter group is
cyclic, so there is a primitive root modulo 14. Note (14) = 6. A quick
check shows that 3 is a primitive root modulo 14. The other primitive
roots are given by 3k (mod 14), where gcd(k, 6) = 1. So, 3, 5(mod 14).
Since x4 1(mod 5) and x2 1(mod 3) for all x with gcd(x, 15) = 1,
we see that x4 1(mod 15) for all such x. But (15) = 8, so there
cannot be primitive roots modulo 15.
2. Note that 5 is relatively prime with 12, the order of (Z/13Z) . Notice
that 5 5 1(mod 12). To solve x5 7(mod 13) raise both sides to
the power 5. We nd, (x5 )5 x25 x 75 11(mod 13). The
solution is x 11(mod 13).
In a similar way we get x5 11(mod 14) x 9(mod 14) and
x5 2(mod 15) x 2(mod 15).
(3.5.35) We use the function (n) from exercise 3.5.14. If we have a primitive root
modulo m then we should have (m) = (m). In other words
lcm(pk11 (p1 1), . . . pkr r (pr 1)) = (pk11 (p1 1), pkr r (pr 1)).
This means that the numbers pki i (pi 1) are all relatively prime. In particular, there can be at most one odd prime factor pi . So m is of the form
m = 2l pk . When k = 0 we have m = 2l and we know that l = 1, 2. When
l = 0 we have m = pk . Suppose that k, l > 0 Then 2l1 and pk1 (p 1) are
relative prime. But this is impossible if l > 1. Hence l = 1 and m = 2pl .
(3.5.29) Case a) We must determine all p such that 10 has order 1,2,3,4,5 or 6 modulo 10. Hence we must determine all p that divide at least one of 101, 102
1, 103 1, 104 1, 105 1, 106 1. This gives us p = 3, 7, 11, 13, 37, 41, 101, 271.
Case b) Using the idea from case a) we get p = 239, 4649.
Case c) p = 73, 137
(3.5.41) From an1 1(mod n) we see that gcd(a, n) = 1. Let k be the order of
a(mod n). Then k|n 1. Suppose that (n 1)/k contains a prime divisor q.
Then a(n1)/q 1(mod n), contradicting our assumptions. Hence (n 1)/k
contaisn no prime divisors, hence (n1)/k = 1 from which we get k = n1.
We also know that k and hence n 1 divide (n). This is only possible if
(n) = n 1, hence n is prime.
134
(3.5.42) Let N = 3 28 + 1. The prime divisors of N 1 are 2, 3. Notice that

11(N 1)/2 1(mod N ) and 11(N 1)/2 360(mod N ). Hence, by Lehmers
test, N is prime.
(5.7.1) Simply write down all squares 12 , 22 , . . . , 82 modulo 17. We get 1, 4, 9, 16, 8, 2, 15, 13(mod 17)
as quadratic residues modulo 17. Similarly we get 1, 4, 9, 16, 6, 17, 11, 7, 5(mod 19)
as quadratic residues modulo 19.
(5.7.2) We rst show that p does not divide b. If it did, then p|a2 + b2 and p|b
would imply p|a, which contradicts gcd(a, b) = 1. Hence p does not divide
b.
Now it follows from p|a2 + b2 that a2 b2 (mod p). Multiply on both sides
by b2 (mod p). We get (ab1 )2 1(mod p). Hence 1 is a quadratic
residue modulo p which implies that p 1(mod 4).
(5.7.4) Part a) Let a be a quadratic nonresidue modulo p and let x = a(p1)/8 .
Then, x4 a(p1)/2 1(mod p).
Part b) Notice that x4 1(mod p) implies that x2 x2 (mod p) and
hence x2 + x2 0(mod p). So we nd that (x + 1/x)2 x2 + 2 + x2
0(mod p). In other words, (x + 1/x)2 2(mod p) and 2 is a quadratic
residue modulo p.
2
(5.7.7) We treat two examples. First
) 114(mod 127). Note that 127 is prime,
( 114 x
so it suces to determine 127 . Notice,
114
127
(
=
2
127
)(
3
127
)(
19
127
The rst factor is 1(because

127
)
( 1 ) 1(mod 8). The second factor, by
127
reciprocity equals ( )3 =
third
( 13 )3 = 1.
( 19 The
)
( 6 )factor, again by reci127
procity,
equals
. The
19 ( )
13
13
( 2 )( 3 )
(191 )
( 114 )latter equals
13
13 13 = (1) 3 = 3 = 1. So we conclude 127 = 1, our
equation is not solvable.
We now study the solvability of 9x2 + 12x + 15 0(mod 58). Splitting o
squares gives (3x + 2)2 + 11 0(mod 58). So it suces to study solvability
of y 2 11(mod 58). By the Chinese remainder theorem this is equivalent
to the system y 2 1(mod 2), y 2( )11(mod 29). The rst equation is
. Note that is equals
solvable, it remains to determine 11
29
(
1
29
)(
11
29
(
=
29
11
(
=
Hence there are no solutions.

7
11
(
=
11
7
( )
4
=
= 1.
7
135
(5.7.8) Let p be an odd prime = 5. Note that
( ) ( )
5
p
=
p
5
The latter is 1 if p 1(mod 5) and 1 if p 2(mod 5).
Now let p be an odd prime = 3. Then
( ) ( )( )
( )
1
3
3
(p1)/2
(p1)/2 p
=
= (1)
(1)
.
p
p
p
3
The latter is 1 if p 1(mod 3) and 1 if p 1(mod 3).
Finally,
( )
( )
3
(p1)/2 p
= (1)
.
p
3
we can now read o that the Legendre symbol is 1 if p 1(mod 12) and
1 if p 5(mod 12).
(5.7.9) Let a = (1)k0 p1 pr be the prime factorisation of a where k0 0 or 1
and the primes pi are not necessarily distinct. Note that the primes pi(are
)
odd because a 1(mod 4). We now compute the Legendre symbol ap
using quadratic reciprocity.
( )k 0
( )
r ( )
1
pi
a
=
p
p
p
i=1
r
r ( )
p
k0 (p1)/2
(pi 1)(p1)/4
= (1)
(1)
pi
i=1
i=1
Now notice the identity
k0 +
(pi 1)/2 (a 1)/2(mod 2)
i=1
On the left we simply have modulo 2 the number of primes pi which are
1(mod 4) and the sign of a. When this total is odd we have a
1(mod 4), when it is even we have a 1(mod 4). So we get
( )
r ( )
a
p
(a1)(p1)/2
.
= (1)
p
p
i
i=1
( p )
only depends on the residue class p(mod |a|),

( )
the value of the sign depends on the parity of (p1)/2. Hence ap depends
The value of the product
pi
136
only on the residue class p(mod 4|a|). Moreover,

( ) if a 1(mod 4) the sign
in front of the product is always +, so now ap depends only on the class
p(mod |a|).
( )
(5.7.10) Note that x2 a(p+1)/2 a(p1)/2 a ap a a(mod p).
(5.7.13) Let a be the smallest quadratic nonresidue and let b the smallest positive
number such that ab > p. Then, 0 < ab p < a. Hence, by minimality of
a, the residue class ab(mod p) must be quadratic residue class. Hence b is
a quadratic non-residue. We also have ab < p + a, hence b < p/a + 1. Since
also a + 1 b it follows that a < p/a and so, a < p.
(5.7.14(a)) Notice that the sum [ p] + + [ kp] equals the number of lattice point,
with positive coordinates, below the graph of px in the interval 1 x k.
Notice that [ pk] = [ p(p 1)/4] = (p 1)/2. The graph of px does

not contain lattice points when 1 x k. To do the counting we might as
well take the number of lattice points in the rectangle 1 x k, 1 y
(p 1)/2 and the subtract the number of lattice points on the right of the
graph. Hence,
(p1)/2 [ 2 ]
y
k(p 1)
.
2
p
l=1
Note that
]
{ 2}
y2
y
y2
=
.
p
p
p
Take the sum for y= 1, 2, . . . , (p 1)/2. The rst part can be summed
using the formula nl=1 n2 = 16 n(n + 1)(2n + 1). The sum of the second
part is precisely equal to K/p where K is the sum of the quadratic residues
modulo p. In case p 1(mod 4) this equals half the sum of all residue
classes, which is p(p 1)/2.
(7.5.1) The sum nX r2 (n) equals the number of lattice points within the disc
with radius X. To every lattice point (a, b) we associate the elementary

square {(a + x, b + y)|0 x, y 1}. Let S be the union of these squares.
The number R(X) is precisely equal
the surface area of S. Not that S
to
lies within
thecircle with radius X + 2. Note also that the circle with
radius X 2 is entirely contained in S. Hence
( X 2)2 R(X) ( X + 2)2 .
from which we can derive that |R(X) X| 2 2X + 2.

(7.5.4) Part a),b) are done simply by trying. Part c) can be done by trying, but
also follows from the next exercise.
137
(7.5.5) We write n = 2k [(3/2)k ]1 as sum of k-th powers. Since n < 3k , it can only
be written as repeated sum with terms 1k , 2k . The most economic way to
do this is to use as many terms 2k as possible. The remainder can then be
written as a sum of ones. Note that [n/2k ] = [(3/2)k ] 1 and the remainder
after division of n by 2k is 2k 1. So we require [n/2k ] 1 terms 2k and
2k 1 terms 1k . Hence g(k) 2k + [(3/2)]k 2.
(7.5.6) From the identity it follows that a number of the form 6m2 can be written
as a sum of 12 fourth powers. We simply write n = a2 +b2 +c2 +d2 (possible
by Lagranges theorem) and use the identity.
From the hint: n = 6N + r and write N as sum of four squares, we deduce
that 6N can be written as the sum of 4 12 = 48 squares. Now choose r
such that 0 r 5 and write r as sum of r terms 14 . We conclude that
g(4) 48 + 5 = 53.
To get a renement, note that r need not be chosen between 0 and 5. We
can also choose among the remainders 0, 1, 2, 33, 16, 17, each which is the
sum of at most two fourth powers. Hence g(4) 48 + 2 = 50.
(9.8.1) We can assume that a, b, c form a Pythagorean triple. Suppose, without loss
of generality, that b is even. Then there exist r, s such that a = r2 s2 , b =
2rs, c = r2 + s2 . Hence abc = 2rs(r4 s4 ). If r or s is divisible by 5
we are done. If r, s are not divisible by 5 we have r4 1(mod 5) and
s4 1(mod 5). Hence r4 s4 1 1 0(mod 5). So r4 s4 is divisible
by 5.
(9.8.2) Note that x, y, z 2 is a Pythagorean triple. Assume that y is even, the
case x even being similar. Then there exist r, s N, with gcd(r, s) = 1
and distinct parity, such that x = r2 s2 , y = 2rs, z 2 = r2 + s2 . Note
that r, s, z is again a Pythagorean triple. Now suppose r is even. Then
there exist integers p, q, with gcd(p, q) = 1 and distinct parity, such that
r = 2pq, s = p2 q 2 , z = p2 + q 2 . So we conclude that
x = (2pq)2 (p2 q 2 )2 = p4 + 6p2 q 2 q 4 ,
y = 4pq(p4 q 4 ),
z = p2 + q 2 .
Similarly, when s is even we conclude s = 2pq, r = p2 q 2 , z = p2 + q 2 and

hence
x = (p2 q 2 )2 (2pq)2 = p4 6p2 q 2 + q 4 ,
y = 4pq(p2 q 2 ),
z = p2 + q 2 .
The remaining solutions arise from the previous ones by interchanging x

and y.
(9.8.3) From the previous problem we know that there exist integers p, q such that
abc = 4pq(p2 q 2 )(p2 + q 2 )(p4 + p2 q 2 + q 4 7p2 q 2 ). Modulo 7 this equals
138

4pq(p2 q 2 )(p6 q 6 ). When 7 divides p or q we are done. When 7 does not
divide pq we have p6 q 6 1(mod 7). Hence p6 q 6 1 1 0(mod 7),
and we are done again.
(9.8.4) Factorisation of x2 + y 2 = z 3 yields (x + iy)(x iy) = z 3 . We should

now nd the greatest common divisor d of x + iy and x iy. Note that d
divides the sum 2x and the dierence 2iy. Since x, y are relatively prime
we conclude that d divides 2. Hence, up to units in Z[i], d = 1, 1 + i, 2.
Suppose 2 divides x + iy. This implies that x, y are both even, which is
excluded by gcd(x, y) = 1. Notice that 1 + i divides x + iy if and only
if x, y have the same parity, i.e. they are both odd. But then we have
x2 + y 2 1 + 1 2(mod 4). In other words, x2 + y 2 contains only one
factor 2, so it can never be a square.
We conclude that x + iy and x iy are relatively prime and hence x + iy
is, up to units, a cube in Z[i]. So there exist a, b Z such that
x + iy = (a + bi)3
where = 1, i. Note that each of these units is a cube, so we can
absorb the unit into the cube part. Hence there exist integers a, b such
that x + iy = (a + bi)3 . After comparison of real and imaginary part we
obtain
x = a3 3ab2 ,
y = 3a2 b b3 .
(9.8.6) We start with the equation 2k 3l = 1. Consider the equation modulo 3.
We see that 2k 1(mod 3), hence k should be even. From the equation it
follows that 2k 1 = 3l , hence (2k/2 1)(2k/2 + 1) = 3l . Hence the factors
2k/2 1 are either 1 or a power of 3. Write these factors as 3a , 3b with
b < a and note that their dierence is 2. I.e. 3a 3b = 2. In other words,
(3ab 1) 3b = 2 and we conclude that b = 0 and a b = 1. This implies
that l = a + b = 1 and k = 2. So there are no solutions k, l 2 in this case.
Now we solve 2k 3l = 1. Consider the equation modulo 4. We nd that
3l 1(mod 4). Hence l is even and we can proceed in a similar way
as above. We get 2k = 3l 1 = (3l/2 1)(3l/2 + 1). The factors 3l/2 1
are either 1 or a power of 2. Write the factors as 2a , 2b with b < a. The
dierence is 2, so we get 2a 2b = 2. Hence (2ab 1)2b = 2, from which
we conclude that b = 1 and a b = 1. So k = a + b = 3 and l = 2. This is
the only solution.
NOTE: Catalans conjecture has been solved in 2002 by Michailovich.
(9.8.8) Write 4y 2 = x3 + 1 as 4y 2 1 = x3 . Factor the left hand side, (2y 1)(2y +
1) = x3 . The numbers 2y 1, 2y + 1 are odd and have dierence 2. So
they are relatively prime and from (2y + 1)(2y 1) = x3 it follows that
139
2y + 1 = u3 and 2y 1 = v 3 are cubes. Hence u3 v 3 = 2. The dierence
of two cubes can only be 2 when u = 1, v = 1. One way to see this is to
note that (u v)(u2 + uv + v 2 ) = 2. Hence u2 + uv + v 2 = 1, 2. The
solutions are then found by trying.
So the nal solution is y = 0, x = 1.
(9.8.10) We prove the rst statement by induction on n. For n = 1 we note that 8
n
n1
n1
divides 32 1. For larger n we remark that 32 1 = (32 1)(32 + 1)
n1
and use the induction hypothesis 2n1 divides 32
1 and the fact that
the second factor is even.
We take ck = 32 , a = 1, b = ck 1. From the above remark we know that
2k+2 divides bk . Note that
k
N (ak bk ck ) 3N (bk )
3
b
<
ck .
k
2k+1
2k+1
Hence ck /N (ak , bk , ck ) > 2k+1 /3. The latter tends to as k .

(9.8.13) To show this we assume that x3 > y 2 and dene = x3 y 2 and We
then apply the abc-conjecture to a = /d, b = y 2 /d, c = x3 /d where d =
gcd(x3 , y 2 ). For every > 0 we get
x3 /d < c()rad(x3 y 2 /d3 )1+ .
Notice that rad(x3 y 2 /d3 ) xy/d. Also notice that y <
So we get
x3 /d < c()(xy/d)1+ < c()(x5/2 /d)1+ .
x3 < x3/2 .
After multiplication by d1+ we obtain

x3 x3 d < c()(x5/2 )1+ .
Now choose such that 5/2 + 3/(1 + ) = a. Then it follows that
xa < c()1/(1+)
from which our assertion follows.
When y 2 > x3 we put = y 2 x3 and nd, as above, that
y 2a/3 < c(a).
Noticing that, by assumption, x < y 2/3 , and our assertion follows also in
this case.
140
(10.3.1) To show part (a) we expand each factor in the product as a geometric series
1
1
1
1
=
1
+
+
+
+
1 ps
ps p2s p3s
Taking the product we see that
pX,p
1
1 ps
prime
equals the sum of n1s over all n which consist entirely of primes X. This
is certainly larger than the sum of n1s over all n X.
We should actually have X > 3 in part (b). To show part (b) one uses the
integral criterion
X1 dt
>
= log(X 1)
t
1
nX
With a bit more case we can also get the lower bound log(X):
X1
dt
>1+
= 1 log(2) + log(X 1) > log(X).
t
2
nX
From (a) and (b) with s = 1 it follows that
pX,p
Take logs on both sides
1
> log(X).
1
1
p
prime
log(1 p1 ) > log log(X).
pX,p prime
Some calculus shows that log(1 x) < x + 4x2 /5 for all x [0, 1/2].
Hence
4
1
+ 2 > log log(X).
p 5p
pX,p prime
Notice also that the sum of 1/p2 over all primes p can be bounded above
by the sum of 1/n2 over all integers n = 1, 4. The latter sum equals 2 /6
1 1/16 = 0.58.... Times 4/5 this yields a number < 1/2. So we get
1
1
+
> log log(X)
2 pX,p prime p
as desired.
141
(10.3.4) Notice that
p2 + 1
=
21
p
p prime
=
p
p2 1
(p2 1)2
prime
1 p4
(1 p2 )2
prime
(2)2
( 2 /6)2
5
= 4
= .
(4)
/90
2
(10.3.5) Notice that, by denition, n = (pn ). Hence, by the prime number theorem,
1 pn
= 1.
n n log(pn )
lim
It now remains to show that

log(pn )
= 1.
n log(n)
lim
This follows from the fact that for suciently large n we have
1 pn
pn
<n<2
,
2 log(pn )
log(pn )
which implies
log(pn ) log log(pn ) log(2) < log(n) < log(pn ) log log(pn ) + log(2).
After division by log(n) and letting n we nd the desired limit.
(10.3.7) From the previous exercise we know that pn , the n-th prime is asymptotic
to n log(n). In particular, for suciently large n, pn > n log(n)/2 > n.

So we get
1
2
< f inite part +
.
p
log(p)
n
log(n)
log(n)/2
prime
n large
The latter innite series converges by the integral criterion.

Similarly we have for suciently large n that pn < 2n log(n) < n2 . Hence
log(p)
2 log(n)
> f inite part +
.
p
n log(n)/2
p prime
n large
The latter innite series is the harmonic series which diverges.
142
(11.4.3) Suppose that e+ and e are rational. Then the polynomial (X e)(X )
has rational coecients. Hence its zeros e, would be quadratic numbers,
contradicting the fact that e is transcendental.
(11.4.4) Suppose the series has a rational value, say p/q. Choose k and consider the
dierence
( )n
p
4
1
k
=
n=0
.
q
5
3n2
2
This is a non-zero rational number with denominator dividing q5k 3k . So

2
we get that 1q 5k 3k . On the other hand we have
( )n
4
1
=
.
5
3n2
n=k+1
Let us estimate the terms of this series by (4/5)n 3(k+1) . Hence

2
<
n (k+1)2
(4/5) 3
(k+1)2
<3
2
(4/5)n = 5 3(k+1) .
n=0
n=k+1
Comparison of the bounds show that

1 k k2
2
5 3
< 5 3(k+1) .
q
Multiplication by 3k gives us 1q 5k < 5 32k1 , hence (9/5)k < 5q/3. This
is impossible if we choose k big enough. Hence our number is irrational.
2
Chapter 13
Appendix: Elementary algebra
13.1
Finite abelian groups
In this section we consider some elementary facts on nite abelian groups. The
main application in this course will be to groups of the form (Z/mZ) , the invertible residue classes modulo m. So, if one does not like to work with general
groups one should simply read (Z/mZ) whenever the expression nite abelian
group is used. The order of a nite group is simply the number of elements it
contains.
Lemma 13.1.1 Let G be a nite abelian group of order |G|. Then a|G| = e for
any a G.
Proof. Let a G. Consider the product P = gG g. Notice that if g runs

through G, then so does ag. Hence
P =
g=
ag.
gG
gG
The latter product can also be written as a|G|

P = a|G| P and hence e = a|G| .
gG
g, which equals a|G| P . So,

2
Remark 13.1.2 Lemma 13.1.1 is actually true for any nite group, also the nonabelian ones. The proof we gave here, due to Lagrange, works only for abelian
groups.
Definition 13.1.3 Let G be a nite group and g G. The smallest positive
integer k such that g k = e is called the order of g. Notation: ord(g).
Lemma 13.1.4 Let G be a nite group and g G. Suppose g k = e. Then
ord(g)|k.
143
144
CHAPTER 13. APPENDIX: ELEMENTARY ALGEBRA
Proof. From g k = e follows g kqord(g) = e for any q Z. Choose q such that 0

k qord(g) < ord(g) and write r = k qord(g). so g r = e. Since 0 r < ord(g)
the minimality of ord(g) implies that r = 0 and thus that ord(g) divides k. 2
Corollary 13.1.5 Let G be a nite abelian group. Then ord(g) divides |G| for
any g G.
Lemma 13.1.6 Let G be a nite abelian group and g, h G. If ord(g) and
ord(h) are relatively prime then ord(gh) = ord(g)ord(h).
Proof. Let M = ord(gh). From e = (gh)M it follows that e = (gh)M ord(g) =
hM ord(g) . Hence ord(h)|M ord(g). Since (ord(g), ord(h)) = 1 we conclude ord(h)|M .
Similarly, ord(g)|M . Hence ord(g)ord(h)|M . On the other hand, (gh)ord(h)ord(g) =
e and so M |ord(g)ord(h) and thus we nd that M = ord(g)ord(h).
2
Lemma 13.1.7 Let G be a nite abelian group and g, h G. Let L be the lowest
common multiple of the orders ord(g), ord(h). Then there exists k G such that
L = ord(k).
Proof. Write L as a product of L1 and L2 such that gcd(L1 , L2 ) = 1 and
L1 |ord(g), L2 |ord(h). Notice that g ord(g)/L1 has order L1 . Similarly hord(h)/L2
has order L2 . According to our previous Lemma the product g ord(g)/L1 hord(h)/L2
has order L1 L2 = L.
By repeated application of this Lemma to more than two elements of G we obtain
the following result.
Corollary 13.1.8 Let G be a nite abelian group and let L be the lowest common
multiple of the orders of all elements of G. Then there exists an element in G of
order L.
Definition 13.1.9 Let G be nite group. The annihilator of G is the smallest
positive integer k such that g k = e for all g G. Notation: Ann(G).
Lemma 13.1.10 Let G be a nite abelian group. Then, Ann(G) = |G|
G is cyclic.
Proof. The being trivial (why?) we shall assume that Ann(G) = |G| and
prove that G is cyclic.
Let L be the lowest common multiple of the orders of all g G. Notice that
Ann(G) L |G|.
The equality Ann(G) = |G| implies |G| = L. According to Lemma 13.1.7 there
exists g G such that ord(g) = L. Hence, by the equality L = |G|, the element
g generates G and thus G is cyclic.
2
13.1. FINITE ABELIAN GROUPS
145
Lemma 13.1.11 Let R be a commutative domain. Then any nite subgroup of

R (the unit group of R) is cyclic.
Proof. Suppose G R is a nite subgroup. Let k = Ann(G) and suppose
that |G| > k. Choose g1 , . . . , gk+1 G distinct. Consider the VanderMonde
determinant

1 1 ...
1

g1 g2 . . . gk+1

= ..
..
.
.

gk gk . . . gk
1
2
k+1
On the one hand this determinant equals i<j (gi gj ). On the other hand,
since gik = 1 i, we have = 0. Since R has no zero divisors this implies that
gi = gj for some i = j, contradicting our choice of distinct elements. So we must
assume that k = |G| and, by Lemma 13.1.10 , G is cyclic.
2
As we said before, the main application of this section is to groups of the form
(Z/mZ) . In particular, Z/pZ is a eld when p is prime. Lemma 13.1.11 implies
that (Z/pZ) is a cyclic group. This can be illustrated by the following example
modulo 13. We have
(Z/13Z) = {1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12}
A possible generator of (Z/13Z) is the element 2. Indeed,
(21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 210 , 211 , 212 ) (2, 4, 8, 3, 6, 12, 11, 9, 5, 10, 7, 1)
and we see that every invertible residue class modulo 13 is a power of 2 modulo
13.
Lemma 13.1.12 Let R be a domain and G a nite subgroup of R . Then,
{
1 if 1 G
g=
1 if 1 G
gG
Proof. Notice that x = 1/x x2 = 1 for any x R. Since R is a domain the

only solutions of x2 = 1 are x = 1. So, in the product
g
gG
all elements cancel except g = 1 and g = 1 if 1 G. The product of these

exceptional elements is 1 if 1 G and 1 if 1 G, which proves our lemma. 2
This lemma, applied to R = Z/pZ for any prime p, yields the following result,
(p 1)! 1(mod p). This is known as Wilsons theorem.
146
13.2
Euclidean domains
Let R be a (not necessarily commutative) domain. The reason to consider non

commutative domains as well is that we would like to include the quaternions in
our considerations. It will not make our statements more dicult, except that we
have to be careful to speak about left- and right- versions of concepts such as
divisibility. For commutative domains we can drop the suxes left- and right-.
Definition 13.2.1 A domain R is called (right) euclidean if there exists a function g : R {0} 7 N such that
i. g(ab) g(b), a, b R {0}
ii. To every a, b R with b = 0 there exist q, r R such that
a = bq + r,
r = 0 or g(r) < g(b).
Definition 13.2.2 Let R be a domain and a, b R. Then b is called a (right)

divisor of a if there exists c R such that a = cb.
Theorem 13.2.3 Let R be a (right) euclidean domain and suppose that a1 , . . . , an
R are not all zero. Then there exists a common (right) divisor d R of a1 . . . , an
such that d = ti ai + tn an for suitable t1 , . . . , tn R.
Proof. Choose d such that g(d) = min{g(x)| x S {0}}, where
S = {x1 a1 + + xn an | x1 , . . . , xn R}.
Then d is a (right) divisor for any a R. To see this choose q R such that
a qd is either 0 or g(a qd) < g(d). Since a qd S the latter inequality would
contradict the minimality of d. Hence a qd = 0. In particular d is common
(right) divisor of all ai . Furthermore we know that d S, so d = t1 a1 + + tn an
for suitable t1 , . . . , tn .
2
As a bonus it follows from Theorem 13.2.3 that any common (right) divisor of the
ai is also a divisor of d. Moreover, by property (i) for euclidean domains we have
g(d) g(d ) for any (right) divisor d of d. So d can truly be called a greatest
common divisor of a1 , . . . , an .
Lemma 13.2.4 Let R be a (right) euclidean domain. Suppose g(1) = 1. Then,
R g() = 1.
Proof. Choose such that = 1. Via property (i) of euclidean domains
we see that g(1) g(). Since g(1) = 1 and g() N this implies g() = 1.
According to property (ii.) of euclidean domains there exists q R such that
either 1 q = 0 or g(1 q) < g(). The latter inequality is impossible, hence
we have 1 = q hence R .
2
The simplest example of a euclidean domain, Z, is treated in a separate chapter.
In the next three sections we shall deal with three examples which are also of
interest in number theory.
13.3. GAUSSIAN INTEGERS
13.3
147
Gaussian integers
Definition 13.3.1 The ring of Gaussian integers is the subring of C given by

{a + bi| a, b Z}. Notation: Z[i]. The norm of an element = a + bi Z[i] is
given by N := a2 + b2 .
Theorem 13.3.2 We have
a) N = N N for any , Z[i]
b) Z[i] is a euclidean domain with the function g() = N
c) The following statements are equivalent,
i. is a unit in Z[i]
ii. N = 1
iii. {1, 1, i, i}.
Proof. a) Note that N = ||2 where | | is the ordinary absolute value on C.
Our assertion follows from || = || ||.
b) Notice that N = 0 = 0 and N 1 for all = 0. Property (i) for
euclidean domains now follows from a). To show property (ii) we work momentarily in C. Let / = x + yi C. Let = p + qi where p, q are the nearest
integers to x and y respectively. So,
|(x + yi )|2 (1/2)2 + (1/2)2 = 1/2.
Hence, after multiplication by N , N ( ) N /2. This proves property
(ii) for euclidean domains.
c) Equivalence of i. and ii. follows from Lemma 13.2.4 and the fact that N 1 = 1.
Equivalence of ii. and iii. follows from the observation
a2 + b2 = 1 (a, b) {(1, 0), (1, 0), (0, 1), (0, 1)}.
2
13.4
Quaternion integers
First of all we recall the denition of quaternions.

Definition 13.4.1 The quaternions are expressions of the form a + bi + cj +
dk, a, b, c, d R with addition given by
(a + bi + cj + dk) + (a + b i + c j + d k) = (a + a ) + (b + b )i + (c + c )j + (d + d )k
148
and multiplication generated by the rules

i2
ij
jk
ki
= j2 =
= k,
= i,
= j,
k2
ji
kj
ik
= 1,
= k,
= i,
= j.
Notation: H
With the above addition and multiplication the quaternions form a (non-commutative)
domain. The conjugate of a quaternion = a+bi+cj+dk is given by abicjdk
and denoted by . The norm of a quaternion = a + bi + cj + dk is given by
N = and equals a2 + b2 + c2 + d2 (verify!). This implies that the inverse of
a non-zero quaternion exists and is given by /N . Furthermore,
Theorem 13.4.2 Let and be quaternions. Then N = N N .
Proof. Write = a + bi + cj + dk and = a + b i + c j + d k. Then,
(a + bi + cj + dk)(a + b i + c j + d k) =
= (aa bb cc dd )
+(ab + ba + cd dc )i
+(ac bd + ca + db )j
+(ad + bc cb + da )k
Our statement now follows from Eulers identity,
(a2 + b2 + c2 + d2 )(a2 + b2 + c2 + d2 ) =
= (aa bb cc dd )2
+(ab + ba + cd dc )2
+(ac bd + ca + db )2
+(ad + bc cb + da )2 .
2
An alternative way to describe quaternions is to identify a + bi + cj + dk with
the matrix
(
)
a + bi c + di
c + di a bi
where i = 1. Addition and multiplication of quaternions comes down to addition and multiplication of the corresponding matrices. The matrix corresponding
to the conjugate reads
(
)
a bi c di
.
c di a + bi
13.4. QUATERNION INTEGERS
149
The norm is simply the determinant of the corresponding matrix. Associativity

of quaternion multiplication now follows directly from associativity of matrix
multiplication. Also, the relation = a2 + b2 + c2 + d2 is now easy to verify.
Definition 13.4.3 Let notations be as above. The quaternion integers are given
by the set
1
{ (p + qi + rj + sk)| p, q, r, s Z, p q r s(mod 2)}.
2
Notation: q.
We leave it to the reader to verify that q is a subring of H and that N N for
all q, = 0.
Theorem 13.4.4 We have
a) q is a (right-)euclidean domain with the function g() = N .
b) The following statements are equivalent,
i. is a unit in q
ii. N = 1
iii. {1, i, j, k, 12 (1 i j k)}.
c) To any quaternionic integer there exists a unit such that is of the
form p + qi + rj + sk, p, q, r, s Z.
Proof. a) Notice that N = 0 = 0 and N 1 for all = 0. Property
(i) for euclidean domains now follows from Theorem 13.4.2. To show property
(ii) we work momentarily in H. Let 1 = x + yi + zj + uk H. Let =
p + qi + rj + sk where p, q, r, s are the nearest integers to x, y, z, u respectively.
So, N (x + yi + zj + uk ) (1/2)2 + (1/2)2 + (1/2)2 + (1/2)2 = 1. Hence, after
multiplication by N , N ( ) N . This proves property (ii) for euclidean
domains when strict inequality holds. Notice that the equality sign holds if and
only if all four numbers x, y, z, u are halntegral. But then we have automatically
x + yi + zj + uk q. So property (ii) holds in all cases.
b) Equivalence of i. and ii. follows from Lemma 13.2.4 and the fact that N 1 = 1.
Equivalence of ii. and iii. follows by determination of all integral solutions of
a2 + b2 + c2 + d2 = 4, a b c d(mod 2).
c) If is already of the form p + qi + rj + sk, p, q, r, s Z we can take = 1. If
is of the form (p + qi + rj + sk)/2, p, q, r, s Zodd we write p = 4p1 + p2 , q =
4q1 + q2 , r = 4r1 + r2 , s = 4s1 + s2 with p2 , q2 , r2 , s2 {1}. Then
= 2(p1 + q1 i + r1 j + s1 k) + (p2 + q2 i + r2 j + s2 k)/2.
Observe that we can now take = (p2 q2 i r2 j s2 k)/2
2
It is a nice exercise to show that the unit group in q modulo 1 is isomorphic to
the alternating group A4 .
150
13.5
Polynomials
Let F be a eld and F [X] the corresponding polynomial ring in one variable. As
well known, any non-zero polynomial f F [X] has a degree, which we denote by
deg(f ). We have deg(f g) = deg(f ) + deg(g) and deg(f + g) deg(f ) + deg(g).
Theorem 13.5.1 Let F [X] be a polynomial ring over a eld F . Then,
a) F [X] is a euclidean domain with the function g(f ) = 2deg(f ) if f 0.
b) f F [X] if and only if f is a constant non-zero polynomial.
Proof. Property (a) follows from the well-known division algorithm for polynomials. Property (b) is a consequence of Lemma 13.2.4
Index
amicable numbers, 17
Bertrands postulate, 110
Cantor diogonal, 121
Carmichael numbers, 34
character, 62
circle method, 78
conductor, 100
congruent, 21
congruent number, 95
conjugate, 85
continued fraction algorithm, 82
convergents, 83
critical strip, 109
decryption, 44
descent, 97
diophantine equation, 94
Dirichlet character, 64
discriminant, 85
divisor sum, 15
encryption, 44
euclidean algorithm, 9
Euler product, 109
even character, 65
Fermat numbers, 37, 54
Gauss sum, 65
greatest common divisor, 7
Hilbert tenth problem, 94
invertible, 21
irrational, 117
Jacobi sum, 66
Jacobi symbol, 56
Jacobsthal sum, 69
lcm, 11
Legendre symbol, 47
lowest common multiple, 11
Mersenne numbers, 17, 37
Mersenne prime, 17
minimal period, 28
minimal polynomial, 85
Mordell equation, 99
negative resdue class, 49
non-residue, 47
odd character, 65
partial fractions, 83
perfect number, 17
period, 28
periodic, 28
periodic continued fraction, 86
Pollard rho, 40
positive residue class, 49
primitive root, 25
principal character, 62
public key, 44
purely periodic, 28, 86
Pythagorean triplet, 95
quadratic
quadratic
quadratic
quadratic
irrational, 85
non-residue, 47
reciprocity, 48
residue, 47
radical, 100
reciprocity law, 67
151
152
reduced quadratic irrational, 85
relatively prime, 7
repunits, 53
residue class, 21
Riemann hypothesis, 109
secret key, 44
sieve methods, 111
terminate, 82
totient function, 24
transcendental, 120
trivial character, 62
twin prime, 110
Warings problem, 78
witness, 35
zer-knowledge proofs, 45
INDEX

Elementary Number Theory - Getalscript

Uploaded by

Copyright:

Available Formats

Elementary Number Theory - Getalscript

Uploaded by

Document Information

Original Description:

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

Elementary Number Theory - Getalscript

Uploaded by

Copyright:

Available Formats

What are the main topics covered in the document?

What are the main topics covered in the document?

What is the Euclidean algorithm and what is it used for?

What is the Euclidean algorithm and what is it used for?

ELEMENTARY NUMBER THEORY

4 Primality and factorisation

6 Dirichlet characters and Gauss sums

12 Solutions to selected problems

13 Appendix: Elementary algebra

F.Beukers, Elementary Number Theory

Roughly speaking, number theory is the mathematics of the integers. In any

2. For all a, b, c N: (a + b) + c = a + (b + c) and (ab)c = a(bc) (associativity

CHAPTER 1. INTEGERS AND THE EUCLIDEAN ALGORITHM

1.2. GREATEST COMMON DIVISORS

Greatest common divisors

CHAPTER 1. INTEGERS AND THE EUCLIDEAN ALGORITHM

viii. gcd(a, b) = 1 there exist x, y Z : ax + by = 1.

Euclidean algorithm for Z

gcd(654321, 123456) = gcd(654321 5 123456, 123456) = gcd(37041, 123456)

F.Beukers, Elementary Number Theory

1.3. EUCLIDEAN ALGORITHM FOR Z

654321 = 5 123456 + 37041

yi+1 = yi1 [ri1 /ri ] yi

Then we can show by induction on i that ri = axi + byi for all i 1. In

CHAPTER 1. INTEGERS AND THE EUCLIDEAN ALGORITHM

In other words, the number of steps in the euclidean algorithm is bounded by a

Fundamental theorem of arithmetic

1.4. FUNDAMENTAL THEOREM OF ARITHMETIC

where p1 , . . . , pr are distinct primes and 0 ki , mj (i, j = 1, . . . r). Then,

CHAPTER 1. INTEGERS AND THE EUCLIDEAN ALGORITHM

Exercise 1.5.1 Prove, n N :

Exercise 1.5.2 Prove, 2n 1 is prime n is prime.

where t runs through Z.

b) Prove that k < log a/ log 2.

c) Prove by induction on n = k, k 1, . . . , 1 that rn ( 12 (1 + 5))kn .

d) Prove that k < log a/ log( 12 (1 + 5)).

CHAPTER 1. INTEGERS AND THE EUCLIDEAN ALGORITHM

Exercise 1.5.18 Let b, c N and suppose (b, c) = 1. Prove,

F.Beukers, Elementary Number Theory

In particular we write = 1 , the sum of divisor and d = 0 , the number

CHAPTER 2. ARITHMETIC FUNCTIONS

i. l (n) is a multiplicative function.

ii. Let n = pk11 pkr r . Then

A very ancient problem is that of perfect numbers.

2.1. DEFINITIONS, EXAMPLES

Definition 2.1.4 A perfect number is a number n N which is equal to the sum

CHAPTER 2. ARITHMETIC FUNCTIONS

Definition 2.2.1 Let f and g be two arithmetic functions. Their convolution

It is an easy excercise to verify that the convolution product is commutative and

Then, for any n N,

Proof. More cryptically we have F = E f and we must prove that f = F .

(d) = (1) + (p) + + (pk )

CHAPTER 2. ARITHMETIC FUNCTIONS

F.Beukers, Elementary Number Theory

CHAPTER 3. RESIDUE CLASSES

Chinese remainder theorem

Theorem 3.2.1 Let m, n Z and (m, n) = 1. Then the natural map :

3.2. CHINESE REMAINDER THEOREM